memory · plan

tl;dr

Memory is not a fact-candidate factory. It is its own layer. Only claims that survive long enough — with reinforcement, and only on ecological topics — may later be proposed manually to the verification pipeline. This page is the map.

the shape

Episode → Memory — lives here, gets reinforced or decays, most just fade.
↓ only if it survives long enough
Fact candidate → real verification pipeline (create_candidate_fact → submit_verification) → Fact.

Memory is mutable, drifting, personal. Fact is hash-sealed, multi-pass verified, load-bearing. They live in different worlds and speak through a narrow, manual bridge.

phases

Phase 1 — separate memory from fact shipped 2026-07-28

Added Mem_Config { BRIDGE_ENABLED: false }.
Gated nominate() and approve_candidate() with BridgeRetired.
1004 existing candidates preserved as historical data; nothing deleted.
Rollback: flip BRIDGE_ENABLED to true. One click, no code change.

Phase 2 — wire the middle hop shipped 2026-07-28

The harvester now extracts :Mem_Fact from every episode via qwen2.5:14b on local Ollama. Enforce mode is on, timer runs hourly, ~2,100 facts. All 10 agents have ECOAGENT_MEMORY_PROVIDER=neo4j set; Neo4jMemoryProvider is live and every narrator run prepends a memory block to its user message.

Cafe Discourse ingestion — live (harvester timer).
Recall wiring — 10/10 agents.
Hermes ingestion — deferred; still returns nothing from that source.

Phase 3 — trigger + confidence lever shipped 2026-07-28

Memory is now a living scalar (Prism-style). Live components:

Reinforce: every successful recall() hit fires an asymptotic impulse (Δ = α × (ceiling − current), α=0.15, ceiling=1.0). Daily-idempotent per (mem_fact_id, source, day) via :Mem_ReinforceEvent MERGE.
Bi-temporal: every :Mem_Fact now carries valid_from, invalidated_at, last_reinforced_at, reinforce_count. Live queries filter invalidated_at IS NULL. Supersede-not-overwrite.
Decay sweep: ecoagent-memory-decay.timer runs daily at 03:00 UTC. Exponential decay with half-life 4 weeks, 3-day grace, floor 0.1. Below 0.2 → soft archive (never deleted). Facts with consensus_score ≥ 0.75 are skipped (external corroboration overrides internal-use signal).
Gate: decay stays OFF until DECAY_ENABLED is flipped on (safe default — no memories rot until you say so).

Phase 4 — ecology-scoped manual promotion next

Only ripe memories (survived decay + reinforced N times) on ecological topics may be proposed. Not automatic. Not LLM-driven. A session-level human action.

Ripe = alive in memory ≥ N days AND reinforce_count ≥ K.
Ecology filter: predicate/subject must resolve to an ecological entity (species, watercourse, biome…).
On push, the claim enters the real verification pipeline from the top — no shortcut.

modules borrowed from

Prism (ProsusAI) — reinforce/decay math, push vs pull channel split.
Graphiti (getzep) — bi-temporal validity, superseded-not-deleted contradiction model.
Letta (fka MemGPT) — the three-scope split (episodic / semantic / procedural).
Mem0 — single-pass ADD-only extraction pattern (opposite of our current heavy triage).

full document

The complete plan lives in the repo as eco-memory-triggers-and-confidence-plan.md. A copy is served at:

download the .md