[EPIC][semantic-memory] Durable general semantic memory: arbitrary knowledge → append-only symbolic world model

nsaspy commented

2026-09-08 01:57:03 +00:00

Owner

Depends on: #1
Upstream compiler: lost-rob0t/prolog-rlm#388
Upstream semantic IR: lost-rob0t/prolog-rlm#392

Mission

Make symbolic-memory the durable long-term consumer of Prolog-RLM's general semantic knowledge compiler.

The goal is not merely to remember facts or how-to procedures. The store must be capable of retaining essentially any semantic knowledge expressible through the upstream extensible knowledge algebra:

exact remembered source
        ↓
prolog-rlm general semantic compiler
        ↓
validated general semantic IR
        ↓
symbolic-memory projection transaction
        ↓
append-only semantic knowledge history
        ↓
current namespace/world-model views
        ↓
Prolog query / inference / retrieval / expert reasoning

Knowledge scope

Symbolic Memory must persist and query all upstream semantic classes, including combinations of them:

entities, identity, aliases, types and taxonomies;
properties and n-ary relations;
definitions and concept equivalence/distinction;
assertions and explicit negative knowledge;
strict/generalized rules;
defaults, exceptions and defeasible rules;
procedures/how-to knowledge, branches, failure and recovery;
events, states and state transitions;
temporal relations and validity intervals;
causal/diagnostic knowledge;
quantitative facts, units and constraints;
spatial relations;
normative/deontic statements such as required/permitted/forbidden/recommended;
goals, plans, intentions, preferences and priorities;
attributed claims, beliefs and denials;
uncertainty, ambiguity and competing interpretations;
comparisons/rankings;
social/organizational roles and responsibilities;
technical/system concepts and invariants;
cases/examples/precedents;
hypotheses, scenarios and counterfactuals;
provenance, trust, scope and derivation meta-knowledge;
future upstream domain extensions using the stable semantic kernel.

Do not flatten these into generic text tags or one undifferentiated triple table if doing so loses modality, scope, time, attribution, rule structure, procedure structure, quantification, uncertainty or provenance.

Universal kernel + open domain vocabulary

Symbolic Memory persists the upstream versioned semantic kernel and domain vocabulary declarations/refs.

It must remain possible to ingest knowledge from new fields—biology, history, engineering, law/policy, personal notes, software, scientific papers, web research, etc.—without redesigning the memory database each time.

Unknown domain concepts are symbolic knowledge, not executable code.

Motivating examples

Procedural

How do I recover X after failure Y?

Recall an ordered procedure with branch conditions and success criteria.

Declarative/taxonomic

What is X? What type of thing is it? What is it part of?

Answer from definitions/type/part-whole knowledge.

Causal/diagnostic

What could cause symptom S and what evidence distinguishes the causes?

Reason over possible causes, diagnostic tests and evidence.

Temporal

Who maintained project P in June, and what changed afterward?

Reason over roles/events/validity intervals rather than returning all historical statements as simultaneously current.

Conflicting research

What do sources say about claim C?

Return attributed supporting/conflicting claims with provenance and unresolved truth status.

Constraints

Given remembered requirements, what assignments satisfy these limits?

Hand the safe upstream constraint representation to trusted Prolog constraint reasoning.

Repository boundary

Prolog-RLM owns

Upstream #388/#392-#396 own:

semantic knowledge algebra / IR;
natural-language semantic compilation;
discourse/coreference/rule/procedure induction;
safe lowering/reasoning semantics;
compiler validation/normalization;
semantic conformance evaluation;
reusable compiler/tool APIs.

symbolic-memory owns

This repository owns:

binding semantic packages to durable memory/source identities;
append-only semantic record/event storage;
namespace/project/global/session scoping;
source and interpretation trust/lifecycle;
semantic supersession/retraction/conflict/corroboration history;
current semantic world-model/materialized views;
indexing/query over every upstream knowledge class;
corpus/backfill ingestion and resumability;
native/MCP memory surfaces;
bounded downstream projections;
persistent cross-memory entity/relation reconciliation.

Dependency direction remains symbolic-memory -> prolog-rlm, never reverse.

Preserve #1 source invariant

Exact natural-language/source memory remains independently durable.

memory_remember(source)
   -> durable source succeeds
semantic projection
   -> may be absent/pending/ready/failed/stale

Projection failure cannot destroy or invalidate the remembered source.

Append-only semantic world model

Do not overwrite old knowledge to manufacture one clean current state.

Represent evolution explicitly:

supersedes(New, Old).
retracts(Retraction, Old, Reason).
conflicts(A, B).
corroborates(A, B).
invalidates(Event, Record, Reason).

Current semantic memory is a materialized/query view over immutable history.

Historical and current knowledge must remain distinguishable through explicit validity/time/lifecycle semantics.

Cross-memory identity and reconciliation

Large corpora will mention the same concepts repeatedly.

Provide bounded reconciliation for:

same entity across memories;
aliases/synonyms;
same event referenced by multiple sources;
same proposition independently corroborated;
incompatible entity resolution candidates;
new domain vocabulary declarations;
local/document-scoped entities that must not be globally merged.

Entity merging must be reversible/provenance-preserving. Do not destroy original compiler identities.

Epistemic semantics

Preserve distinctions such as:

source says P
memory/compiler inferred P
Prolog derived P
P is currently observed
P was historically true
P is hypothetical
P is explicitly false
P is unknown
P is conflicted

A model-extracted fact does not silently inherit user/system authority from the source memory.

Normative remembered knowledge such as forbidden(X) is knowledge about a policy and does not itself become host execution authority.

Query/world-model requirements

The durable layer must support current/history/query surfaces over arbitrary semantic classes rather than APIs specialized only to facts/procedures.

Conceptual forms:

memory_semantic_query(+Query, +Options, -Outcome).
memory_semantic_current(+SubjectOrScope, +Options, -Outcome).
memory_semantic_history(+SubjectOrScope, +Options, -Outcome).
memory_semantic_explain(+RecordOrResult, -Outcome).
memory_semantic_projection(+Goal, +Options, -Outcome).

Specialized convenience calls such as memory_procedure may exist on top of this general query substrate.

Retrieval / bounded projection

Never dump the whole symbolic world model into a model prompt.

For a goal:

Prolog/retrieval
   -> select relevant entities/facts/rules/events/procedures/constraints/etc.
   -> include compact provenance refs
   -> only expand original source when required

This must work even when the durable knowledge base contains millions of semantic records.

Corpus/backfill scope

The ingestion path should handle:

years of LLM/chat logs;
web research archives;
manuals/docs/books/notes;
issue/PR histories;
agent/tool traces;
reports;
mixed structured/unstructured corpora;
existing exact memories missing semantic projections.

The point is to compile material meaning once and reuse it symbolically instead of rereading the source corpus every time.

Child issues

Existing children remain, but all must consume the full general semantic IR, not a procedural-only subset:

#5 compiler adapter + independent projection lifecycle;
#6 append-only semantic ledger/current-history views;
#7 general symbolic recall/query/reasoning;
#8 resumable corpus/backfill ingestion;
#9 native/MCP exposure and downstream projections.

Upstream implementation issues:

prolog-rlm#392 general semantic IR;
prolog-rlm#393 semantic induction;
prolog-rlm#394 safe reasoning/lowering;
prolog-rlm#395 cross-domain conformance;
prolog-rlm#396 library/tool/large-document integration.

North-star acceptance

Ingest a heterogeneous corpus containing operational instructions, technical docs, historical/event records, scientific claims, policy text, quantitative constraints, organization/role descriptions and conflicting web research.

After restart and with original source text absent from model context, demonstrate symbolic queries that:

derive a procedure from procedural rules;
answer a taxonomy/definition query;
chain causal/diagnostic knowledge;
respect temporal validity;
solve a safe remembered constraint problem;
expose an exception to a default rule;
return conflicting attributed claims without collapsing them;
distinguish unknown from explicit false;
explain every result back to semantic records and exact source spans;
return only a bounded relevant projection.

No LLM/provider is required for recall/reasoning once semantic projections exist, except optional conversational rendering or explicitly chosen fallback reasoning.

Non-goals

no duplicate semantic compiler in symbolic-memory;
no single giant hardcoded domain ontology;
no auto-consult of generated Prolog;
no semantic record becoming execution authority merely because it is remembered;
no requirement that every exact memory have a semantic projection;
no loss of source provenance/history to simplify current views.

Depends on: #1 Upstream compiler: lost-rob0t/prolog-rlm#388 Upstream semantic IR: lost-rob0t/prolog-rlm#392 ## Mission Make `symbolic-memory` the durable long-term consumer of Prolog-RLM's **general semantic knowledge compiler**. The goal is not merely to remember facts or `how-to` procedures. The store must be capable of retaining essentially any semantic knowledge expressible through the upstream extensible knowledge algebra: ```text exact remembered source ↓ prolog-rlm general semantic compiler ↓ validated general semantic IR ↓ symbolic-memory projection transaction ↓ append-only semantic knowledge history ↓ current namespace/world-model views ↓ Prolog query / inference / retrieval / expert reasoning ``` ## Knowledge scope Symbolic Memory must persist and query all upstream semantic classes, including combinations of them: - entities, identity, aliases, types and taxonomies; - properties and n-ary relations; - definitions and concept equivalence/distinction; - assertions and explicit negative knowledge; - strict/generalized rules; - defaults, exceptions and defeasible rules; - procedures/how-to knowledge, branches, failure and recovery; - events, states and state transitions; - temporal relations and validity intervals; - causal/diagnostic knowledge; - quantitative facts, units and constraints; - spatial relations; - normative/deontic statements such as required/permitted/forbidden/recommended; - goals, plans, intentions, preferences and priorities; - attributed claims, beliefs and denials; - uncertainty, ambiguity and competing interpretations; - comparisons/rankings; - social/organizational roles and responsibilities; - technical/system concepts and invariants; - cases/examples/precedents; - hypotheses, scenarios and counterfactuals; - provenance, trust, scope and derivation meta-knowledge; - future upstream domain extensions using the stable semantic kernel. Do **not** flatten these into generic text tags or one undifferentiated triple table if doing so loses modality, scope, time, attribution, rule structure, procedure structure, quantification, uncertainty or provenance. ## Universal kernel + open domain vocabulary Symbolic Memory persists the upstream versioned semantic kernel and domain vocabulary declarations/refs. It must remain possible to ingest knowledge from new fields—biology, history, engineering, law/policy, personal notes, software, scientific papers, web research, etc.—without redesigning the memory database each time. Unknown domain concepts are symbolic knowledge, not executable code. ## Motivating examples ### Procedural ```text How do I recover X after failure Y? ``` Recall an ordered procedure with branch conditions and success criteria. ### Declarative/taxonomic ```text What is X? What type of thing is it? What is it part of? ``` Answer from definitions/type/part-whole knowledge. ### Causal/diagnostic ```text What could cause symptom S and what evidence distinguishes the causes? ``` Reason over possible causes, diagnostic tests and evidence. ### Temporal ```text Who maintained project P in June, and what changed afterward? ``` Reason over roles/events/validity intervals rather than returning all historical statements as simultaneously current. ### Conflicting research ```text What do sources say about claim C? ``` Return attributed supporting/conflicting claims with provenance and unresolved truth status. ### Constraints ```text Given remembered requirements, what assignments satisfy these limits? ``` Hand the safe upstream constraint representation to trusted Prolog constraint reasoning. ## Repository boundary ### Prolog-RLM owns Upstream #388/#392-#396 own: - semantic knowledge algebra / IR; - natural-language semantic compilation; - discourse/coreference/rule/procedure induction; - safe lowering/reasoning semantics; - compiler validation/normalization; - semantic conformance evaluation; - reusable compiler/tool APIs. ### symbolic-memory owns This repository owns: - binding semantic packages to durable memory/source identities; - append-only semantic record/event storage; - namespace/project/global/session scoping; - source and interpretation trust/lifecycle; - semantic supersession/retraction/conflict/corroboration history; - current semantic world-model/materialized views; - indexing/query over every upstream knowledge class; - corpus/backfill ingestion and resumability; - native/MCP memory surfaces; - bounded downstream projections; - persistent cross-memory entity/relation reconciliation. Dependency direction remains `symbolic-memory -> prolog-rlm`, never reverse. ## Preserve #1 source invariant Exact natural-language/source memory remains independently durable. ```text memory_remember(source) -> durable source succeeds semantic projection -> may be absent/pending/ready/failed/stale ``` Projection failure cannot destroy or invalidate the remembered source. ## Append-only semantic world model Do not overwrite old knowledge to manufacture one clean current state. Represent evolution explicitly: ```prolog supersedes(New, Old). retracts(Retraction, Old, Reason). conflicts(A, B). corroborates(A, B). invalidates(Event, Record, Reason). ``` Current semantic memory is a materialized/query view over immutable history. Historical and current knowledge must remain distinguishable through explicit validity/time/lifecycle semantics. ## Cross-memory identity and reconciliation Large corpora will mention the same concepts repeatedly. Provide bounded reconciliation for: - same entity across memories; - aliases/synonyms; - same event referenced by multiple sources; - same proposition independently corroborated; - incompatible entity resolution candidates; - new domain vocabulary declarations; - local/document-scoped entities that must not be globally merged. Entity merging must be reversible/provenance-preserving. Do not destroy original compiler identities. ## Epistemic semantics Preserve distinctions such as: ```text source says P memory/compiler inferred P Prolog derived P P is currently observed P was historically true P is hypothetical P is explicitly false P is unknown P is conflicted ``` A model-extracted fact does not silently inherit user/system authority from the source memory. Normative remembered knowledge such as `forbidden(X)` is knowledge *about a policy* and does not itself become host execution authority. ## Query/world-model requirements The durable layer must support current/history/query surfaces over arbitrary semantic classes rather than APIs specialized only to facts/procedures. Conceptual forms: ```prolog memory_semantic_query(+Query, +Options, -Outcome). memory_semantic_current(+SubjectOrScope, +Options, -Outcome). memory_semantic_history(+SubjectOrScope, +Options, -Outcome). memory_semantic_explain(+RecordOrResult, -Outcome). memory_semantic_projection(+Goal, +Options, -Outcome). ``` Specialized convenience calls such as `memory_procedure` may exist on top of this general query substrate. ## Retrieval / bounded projection Never dump the whole symbolic world model into a model prompt. For a goal: ```text Prolog/retrieval -> select relevant entities/facts/rules/events/procedures/constraints/etc. -> include compact provenance refs -> only expand original source when required ``` This must work even when the durable knowledge base contains millions of semantic records. ## Corpus/backfill scope The ingestion path should handle: - years of LLM/chat logs; - web research archives; - manuals/docs/books/notes; - issue/PR histories; - agent/tool traces; - reports; - mixed structured/unstructured corpora; - existing exact memories missing semantic projections. The point is to compile material meaning once and reuse it symbolically instead of rereading the source corpus every time. ## Child issues Existing children remain, but all must consume the **full general semantic IR**, not a procedural-only subset: - #5 compiler adapter + independent projection lifecycle; - #6 append-only semantic ledger/current-history views; - #7 general symbolic recall/query/reasoning; - #8 resumable corpus/backfill ingestion; - #9 native/MCP exposure and downstream projections. Upstream implementation issues: - prolog-rlm#392 general semantic IR; - prolog-rlm#393 semantic induction; - prolog-rlm#394 safe reasoning/lowering; - prolog-rlm#395 cross-domain conformance; - prolog-rlm#396 library/tool/large-document integration. ## North-star acceptance Ingest a heterogeneous corpus containing operational instructions, technical docs, historical/event records, scientific claims, policy text, quantitative constraints, organization/role descriptions and conflicting web research. After restart and with original source text absent from model context, demonstrate symbolic queries that: 1. derive a procedure from procedural rules; 2. answer a taxonomy/definition query; 3. chain causal/diagnostic knowledge; 4. respect temporal validity; 5. solve a safe remembered constraint problem; 6. expose an exception to a default rule; 7. return conflicting attributed claims without collapsing them; 8. distinguish unknown from explicit false; 9. explain every result back to semantic records and exact source spans; 10. return only a bounded relevant projection. No LLM/provider is required for recall/reasoning once semantic projections exist, except optional conversational rendering or explicitly chosen fallback reasoning. ## Non-goals - no duplicate semantic compiler in symbolic-memory; - no single giant hardcoded domain ontology; - no auto-consult of generated Prolog; - no semantic record becoming execution authority merely because it is remembered; - no requirement that every exact memory have a semantic projection; - no loss of source provenance/history to simplify current views.

nsaspy commented

2026-09-08 01:57:10 +00:00

Author

Owner

Child issue map

The semantic-memory implementation set is now decomposed as:

#5 — Prolog-RLM semantic compiler adapter + independent projection lifecycle
#6 — append-only symbolic projection ledger, supersession/conflict/provenance/current views
#7 — symbolic recall/query for remembered facts, rules and procedures without rereading source text
#8 — resumable corpus/backfill ingestion for LLM logs, web research, docs and historical memories
#9 — native/MCP exposure, bounded symbolic projections, provenance/source expansion, and Prolog-RLM integration

Dependency direction remains symbolic-memory -> prolog-rlm#388. #1 stays the source-durability foundation and should not be widened to implement these features before its current RAGE slice is complete.

## Child issue map The semantic-memory implementation set is now decomposed as: - #5 — Prolog-RLM semantic compiler adapter + independent projection lifecycle - #6 — append-only symbolic projection ledger, supersession/conflict/provenance/current views - #7 — symbolic recall/query for remembered facts, rules and procedures without rereading source text - #8 — resumable corpus/backfill ingestion for LLM logs, web research, docs and historical memories - #9 — native/MCP exposure, bounded symbolic projections, provenance/source expansion, and Prolog-RLM integration Dependency direction remains `symbolic-memory -> prolog-rlm#388`. #1 stays the source-durability foundation and should not be widened to implement these features before its current RAGE slice is complete.

nsaspy commented

2026-09-08 01:57:11 +00:00

Author

Owner

Scope extension: add #10 as the dedicated cross-memory identity/ontology-reconciliation child. The semantic-memory program is now explicitly required to preserve and query the full general semantic IR from prolog-rlm#392, including entity/type/taxonomy, rules/defaults/exceptions, procedures, events/time, causality, quantities/constraints, spatial relations, modalities, goals/preferences, attributed claims, uncertainty/hypotheses/scenarios, cases, and provenance/meta-knowledge. Do not reduce implementation to a fact/rule/procedure subset.

nsaspy commented

2026-09-08 01:57:11 +00:00

Author

Owner

Machine Spirit pass prolog-rlm#398 reached one storage-relevant design decision that should constrain this epic: persist abstract proposition identity separately from assertion/context/provenance identity.

Recommended downstream shape:

proposition P = canonical semantic content under IR/vocabulary version
assertion A1 -> P, asserted, source/span/time/context
assertion A2 -> P, claimed_by(S), source/span/time/context
assertion A3 -> P, explicit_denial, source/span/time/context
derivation D1 -> P, rule/premise refs

Implications:

current/history/conflict/supersession operates primarily over assertion/derivation records, not by rewriting proposition payloads;
many sources may reference one proposition without losing independent provenance;
proposition structural fingerprints are not real-world entity same_as; entity/proposition reconciliation remains explicit reversible knowledge (#10);
native n-ary propositions should not be flattened to triples in storage merely for convenience;
materialized frames/object views/current-state views are rebuildable indexes, not the authoritative semantic record;
persist the upstream semantic profile/dialect declaration so reopened knowledge never gains stronger reasoning semantics than the package was validated for.

Full alternatives, sources, complexity and safe-lowering reasoning are preserved in prolog-rlm#398. This comment is only the downstream storage consequence.

Machine Spirit pass prolog-rlm#398 reached one storage-relevant design decision that should constrain this epic: **persist abstract proposition identity separately from assertion/context/provenance identity**. Recommended downstream shape: ```text proposition P = canonical semantic content under IR/vocabulary version assertion A1 -> P, asserted, source/span/time/context assertion A2 -> P, claimed_by(S), source/span/time/context assertion A3 -> P, explicit_denial, source/span/time/context derivation D1 -> P, rule/premise refs ``` Implications: - current/history/conflict/supersession operates primarily over assertion/derivation records, not by rewriting proposition payloads; - many sources may reference one proposition without losing independent provenance; - proposition structural fingerprints are not real-world entity `same_as`; entity/proposition reconciliation remains explicit reversible knowledge (#10); - native n-ary propositions should not be flattened to triples in storage merely for convenience; - materialized frames/object views/current-state views are rebuildable indexes, not the authoritative semantic record; - persist the upstream semantic profile/dialect declaration so reopened knowledge never gains stronger reasoning semantics than the package was validated for. Full alternatives, sources, complexity and safe-lowering reasoning are preserved in prolog-rlm#398. This comment is only the downstream storage consequence.

nsaspy commented

2026-09-08 01:57:11 +00:00

Author

Owner

Machine Spirit #398B materially deepened the upstream semantic IR contract. Downstream durable memory should now plan to preserve first-class context/theory identity, explicit theory bridges/imports, scoped existential witnesses, and justification/derivation dependencies in addition to propositions/assertions/provenance. Source-context assertions must remain distinguishable from promoted/domain-world assertions; a parsed Wikipedia/news/LLM-log statement does not become global truth merely by entering memory. Canonical append-only semantic history should remain separate from rebuildable derived/materialized state. See lost-rob0t/prolog-rlm#398 depth addendum and updated #392.

Machine Spirit #398B materially deepened the upstream semantic IR contract. Downstream durable memory should now plan to preserve first-class **context/theory identity, explicit theory bridges/imports, scoped existential witnesses, and justification/derivation dependencies** in addition to propositions/assertions/provenance. Source-context assertions must remain distinguishable from promoted/domain-world assertions; a parsed Wikipedia/news/LLM-log statement does not become global truth merely by entering memory. Canonical append-only semantic history should remain separate from rebuildable derived/materialized state. See `lost-rob0t/prolog-rlm#398` depth addendum and updated `#392`.

nsaspy commented

2026-09-08 01:57:11 +00:00

Author

Owner

MACHINE-SPIRIT Depth 5C canonical handoff

prolog-rlm#402C is complete with CLFM — Causal Ledger Federation Mesh. The durable-world-model epic should treat federation as composition of independent logical ledgers through typed/versioned links, not as one global log or universal CRDT merge.

Key invariants for #4:

local order != cross-ledger causal order
replica != independent semantic source
receipt/inclusion != truth/warrant
remote issuer != local authority
redaction notice != remote sanitization completion
same hash != same assertion episode

Federated query/replay uses explicit per-ledger frontiers/checkpoints (federation_cut) with link/mapping/admission/redaction/integrity generations and completeness status. Physical federation/storage implementation remains unproven and is deferred to #402D. Refs #5-#10.

### MACHINE-SPIRIT Depth 5C canonical handoff prolog-rlm#402C is complete with **CLFM — Causal Ledger Federation Mesh**. The durable-world-model epic should treat federation as composition of independent logical ledgers through typed/versioned links, not as one global log or universal CRDT merge. Key invariants for #4: ```text local order != cross-ledger causal order replica != independent semantic source receipt/inclusion != truth/warrant remote issuer != local authority redaction notice != remote sanitization completion same hash != same assertion episode ``` Federated query/replay uses explicit per-ledger frontiers/checkpoints (`federation_cut`) with link/mapping/admission/redaction/integrity generations and completeness status. Physical federation/storage implementation remains unproven and is deferred to #402D. Refs #5-#10.

nsaspy commented

2026-09-08 01:57:11 +00:00

Author

Owner

Library-first implementation decomposition

#4 remains the canonical durable general semantic world-model domain epic under new umbrella #11. It is not superseded.

Implementation ownership is now decomposed across subsystem epics #12-#19 and ordered slices #20-#45. The most direct #4 implementation path is:

#22 semantic projection lifecycle
-> #23 append-only full-IR persistence
-> #24 bitemporal current/history views
-> #25 epistemic support/provenance
-> #26 indexes
-> #27 identity/ontology reconciliation
-> #28 general query
-> #29/#30 bounded attention/retrieval state

Large-corpus, learning-history, scaling/federation, transport, and conformance continue through #31-#45. Workers should pick dependency-ready slices, not code directly from this broad epic.

## Library-first implementation decomposition #4 remains the canonical **durable general semantic world-model domain epic** under new umbrella #11. It is not superseded. Implementation ownership is now decomposed across subsystem epics #12-#19 and ordered slices #20-#45. The most direct #4 implementation path is: ```text #22 semantic projection lifecycle -> #23 append-only full-IR persistence -> #24 bitemporal current/history views -> #25 epistemic support/provenance -> #26 indexes -> #27 identity/ontology reconciliation -> #28 general query -> #29/#30 bounded attention/retrieval state ``` Large-corpus, learning-history, scaling/federation, transport, and conformance continue through #31-#45. Workers should pick dependency-ready slices, not code directly from this broad epic.

nsaspy referenced this issue

2026-09-08 01:57:39 +00:00

[EPIC][CORE] Library-first Symbolic Memory v1: durable semantic world model, learning history, federation, and zero-model recall #38

nsaspy referenced this issue

2026-09-08 01:57:40 +00:00

[EPIC][CORE] Library-first Symbolic Memory v1: durable semantic world model, learning history, federation, and zero-model recall #38

nsaspy referenced this issue

2026-09-08 01:59:55 +00:00

[SLICE 27][CORE-H] Symbolic Memory v1 north-star integration: heterogeneous corpus → durable semantics → bounded zero-model reasoning → learned-history replay #4

Rows
Columns

[EPIC][semantic-memory] Durable general semantic memory: arbitrary knowledge → append-only symbolic world model #45

Mission

Knowledge scope

Universal kernel + open domain vocabulary

Motivating examples

Procedural

Declarative/taxonomic

Causal/diagnostic

Temporal

Conflicting research

Constraints

Repository boundary

Prolog-RLM owns

symbolic-memory owns

Preserve #1 source invariant

Append-only semantic world model

Cross-memory identity and reconciliation

Epistemic semantics

Query/world-model requirements

Retrieval / bounded projection

Corpus/backfill scope

Child issues

North-star acceptance

Non-goals

Child issue map

MACHINE-SPIRIT Depth 5C canonical handoff

Library-first implementation decomposition