Memory

Persistent, searchable state that outlives the process. Structured enough for graph traversal; flat enough for grep.

Components shipped today

SQLite + FTS5 lattice (3.A) — internal/memory/SqliteStore. Schema migrations, fire-and-forget worker, lexical search.
Ontological graph (3.B) — entities, relationships, LLM-assisted extractor with dead-letter queue.
Neural recall (3.C) — 2-layer seed selection, CTE traversal, <memory-context> fence injection matching Hermes’s build_memory_context_block.
Semantic fusion (3.D) — Ollama embeddings, cosine recall, and hybrid lexical+semantic seed fusion.
USER.md mirror (3.D.5) — async export of entity/relationship graph to human-readable Markdown. Gormes-original; no upstream equivalent.
Tool audit JSONL (3.E.2) — append-only JSONL from kernel and delegate_task tool execution with timing, outcome, and error fields.
Transcript export (3.E.3) — gormes session export <id> --format=markdown renders SQLite turns, timestamps, and tool calls for operator sharing.
Operator visibility (3.E.4, 3.E.5) — gormes memory status is shipped, and the local insights layer now persists append-only daily usage.jsonl records from telemetry.Snapshot rollups.
GONCHO compatibility seam — internal memory work lives behind the goncho service, while the exported tool surface remains Honcho-compatible (honcho_*).

Phase 3 closeout queue

Shipped visibility spine (3.E.1–3.E.5) — session index mirror, tool audit, transcript export, memory status, and daily insights logging are landed.
last_seen closeout (3.E.6) — shipped: schema v3g backfills relationships.last_seen, repeated relationship observations advance it without rewriting legacy updated_at, and recall attenuation uses COALESCE(NULLIF(last_seen, 0), updated_at).
Cross-chat identity closeout (3.E.7) — shipped: GONCHO identity hierarchy is user_id > chat_id > session_id; internal/persistence/session persists canonical chat-to-user bindings, and internal/memory, internal/goncho, and internal/gonchotools cover same-chat default fencing, opt-in canonical user/source-filtered recall, Honcho-compatible schemas, host mappings, SillyTavern persona/group-chat mapping, deny paths, and operator-readable evidence.
Session lineage + cross-source search closeout (3.E.8) — shipped: source-filtered search spans one canonical user_id across chats inside internal/memory and the internal GONCHO service; parent_session_id, compression-continuation resume, lineage-aware hits, and operator-auditable search evidence are validated.
Goncho/Honcho parity (3.F) — shipped and converged: context representation options, typed search filters, directional peer cards, queue status, summary budgeting, dialectic chat, file import, topology fixtures, operator diagnostics, streaming persistence, [goncho] config, and dream-scheduler intent are validated while public tools remain honcho_*.

Identity + lineage contract

GONCHO identity hierarchy — user_id > chat_id > session_id.
Recall fence — same-chat by default; opt-in cross-chat only when a canonical user_id resolves.
Tool boundary — honcho_search and honcho_context preserve the external Honcho-compatible tool names and now advertise scope / sources while the implementation stays in the internal goncho package.
Lineage rule — parent_session_id is append-only metadata on descendants, not a rewrite of ancestor history.
Implementation plan — docs/superpowers/plans/memory-identity/2026-04-22-gormes-phase3-identity-lineage-plan.md.
Execution plan — docs/superpowers/plans/memory-identity/2026-04-22-gormes-phase3-identity-lineage-execution-plan.md; the identity and lineage sequence it described has now landed. Future memory work should add new small 3.F or Phase 6 rows rather than reopening the shipped cross-chat spine.

Why this is not just “chat logs”

Chat logs are append-only. Memory has schema. You query it, derive from it, inject it back into the context window. The SQLite + FTS5 combination gives you ACID durability and full-text search in a single ~100 KB binary dependency.

Donor pointers

When implementing a new memory or Goncho slice, route through the gormes-references skill (docs/development-skills/gormes-references/SKILL.md) before re-deriving a shape. The most useful Go donors for memory/Goncho work live under references/go-agent-os/engram/:

Memory problem	Donor file
SQLite + FTS5 store with DDL, indexes, migrations	`engram/internal/persistence/store/store.go`
Memory relation/conflict vocabulary (`related`, `conflicts_with`, `supersedes`, `compatible`, `scoped`, `not_conflict`)	`engram/internal/persistence/store/relations.go`
Deterministic serialized MCP write queue (mutex-serialized, cancel-before-start)	`engram/internal/mcp/write_queue.go`
MCP activity/audit logging (audit shape, redaction, append-only file)	`engram/internal/mcp/activity.go`
Bounded token-budget tracker for recall context assembly	`axe/internal/budget/budget.go`
Artifact tracker with sanitized paths (path-traversal guard, append-only registry)	`axe/internal/artifact/tracker.go`

Engram is MIT-licensed and permitted as patterns + adapted code with attribution. Always add a // Adapted from engram/...::Symbol comment on the receiving Gormes file when porting code, and convert types/imports to Gormes-native names so engram_* symbols never leak into Gormes’ public API.

See Phase 3 for the full sub-status.