self-directed-learning/state/PHASE_2_AND_4_REPORT.md

Adaptive Recall Phase 2 + Phase 4 — Final Report

Date: 2026-07-12 Build context: Adaptive Recall for both Hermes and Krystie, sidecar on port 4380, 851 UMP records, 768/851 in Qdrant memories_ump.

---

Phase 2 (Knowledge Graph Channel)

What was built

File	Lines	Purpose
src/graph.js	414	Graph class with load/save/buildFromRecords/addEntity/addEdge/neighbors/searchEntities/topEntities/topEdges/stats
scripts/build_graph.js	82	Standalone script that loads UMP records and builds/persists the graph
state/graph.json	—	Built artifact: 2592 nodes, 111 edges, 719 URNs
state/SCHEMA.md	—	Documentation of actual UMP record schema (corrects compaction summary)
test/test_graph.js	—	10+ unit tests (the graph test file from subagent)

Wiring

src/server.js modified to:

• Import Graph from ./graph.js
• Add lazy-loaded singleton getGraph() that loads from GRAPH_FILE (env-overridable)
• Add runGraphChannel(): extracts entities from query, BFS up to GRAPH_DEPTH hops (default 2), score = 1/hops
• Wire 3rd channel into /recall: Promise.allSettled([runUmpChannel, runVectorChannel, runGraphChannel])
• Update /health to report graph stats ({enabled, loaded, nodes, edges, urns, file}) and bump phase label to 2C
• Default weights now {ump: 1.0, vector: 1.0, graph: 1.0}

Eval results

Channel-contribution eval (scripts/eval_channel_contribution.py):

Avg graph-channel hits in top-5:
  3ch (graph enabled):    0.85 hits/query
  2ch (graph zero-weigh): 0.30 hits/query
  Delta:                  +0.55

Hits in 3ch top-5 that are NOT in 2ch top-5:
  Total unique: 12 (avg 0.60/query)

Latency: 2ch=903ms  3ch=1079ms  Δ=+176ms (graph overhead)

Notable graph rank-#1 wins (queries where graph found a hit that ranked higher than ump+vector):

• "Triangles fuzz harness CI gflags missing" → graph rank #1
• "Krystie OpenClaw migrated Hermes profile" → graph rank #1
• "Triangles test coverage PR keystore V5 soft cap" → graph rank #1
• "Triangles multisig stack walk test audit" → graph rank #3
• "DashCaddy TOTP recovery panel disable bak fallback" → graph rank #3
• "Triangles PoW PoS cutoff nonce zero" → graph rank #4

Zero regressions — 10/20 queries saw graph-channel hits in top-5; the other 10 stayed at the same rank.

---

Phase 4 (Memory Lifecycle Decay)

What was built

File	Purpose
scripts/ump_decay.py	Standalone Python script with pure apply_decay(records, dry_run) function, atomic writes, timestamped backups
test/test_ump_decay.py	20 unit tests across 10 test cases, stdlib-only
Cron eaff2d9683fc	Runs nightly at 3am

Decay rates (per day since last access)

Kind	λ	Notes
identity	0.0001	very slow, almost never decays
semantic	0.001	facts over months
procedural	0.005	medium, skills fade if unused
note	0.003	medium-slow, session notes
episodic	0.01	events over weeks
working	0.05	fast, session context

Formula: confidence_new = confidence_old * exp(-λ * days_since_modified_or_created) Floor at 0.05.

Status transitions:

• candidate → active if confidence >= 0.5
• active → archived if confidence < 0.2
• archived → archived (no resurrection)
• tombstoned skipped entirely

Dry-run on real data (851 records)

Status transitions:
  active    708 → 760  (+52 from candidate promotion)
  candidate 53  → 1    (−52)
  archived  42  → 42
  tombstoned 48 → 48   (skipped)

Archives triggered: 0 (oldest record is 105 days; even working λ=0.05 × 20d only decays ~64%)
Promotions: 52 (mostly procedural/semantic at conf=0.75-0.85 within last 36 days)

Confidence histogram (decayed subset, 803 records):
  [0.20-0.40) = 93
  [0.40-0.60) = 213
  [0.60-0.80) = 286
  [0.80-1.00] = 211

Edge cases found:
  - 809/851 records missing time.modified (defaults to time.created) — handled
  - No weird timestamps, no missing lifecycle blocks
  - float-edge in exp(-λ*0) yields ~1.0 - 3e-13 (handled with tolerance)

Tests

20/20 PASS — all 10 spec test cases plus extras.

---

Sidecar status

$ curl -s http://127.0.0.1:4380/health | python3 -m json.tool
{
  "status": "ok",
  "qdrant": "reachable",
  "ump": "reachable",
  "ollama": "reachable",
  "graph": {
    "enabled": true,
    "loaded": true,
    "nodes": 2592,
    "edges": 111,
    "urns": 719,
    "file": "/root/ump-recall/state/graph.json"
  },
  "qdrant_collection": "memories_ump",
  "upstreams": {...},
  "phase": "2C"
}

Sidecar restarted cleanly (old PID 4036490 killed, new PID 4102494 via background process).

---

Both gateways now use 3-channel RRF

Verified earlier:

• Hermes config patched (via terminal sed, since patch tool was security-guarded)
• Krystie config already pointing at shim
• MCP shim routes recall to sidecar, falls back to canonical UMP if sidecar down
• All other tools pass through to canonical UMP

---

What's still TODO

1. Cron first run — ump-decay-nightly runs 2026-07-13 03:00. First dry-run before apply is the safety check.
2. Refresh graph.json periodically — when UMP gets new records, the graph is stale. Add a cron to run node scripts/build_graph.js weekly.
3. Filter noisy edges — some spurious relations like "hours→last" come from arrow regex matching English. Could add a stopword filter.
4. Refine entities.js — the graph layer added EXTRA_ENTITY_PATTERNS (lowercase services, port numbers, krystie-* codes) because the core extractor only knows a fixed allow-list. Could merge these into the main extractor.
5. Pre-existing MCP shim bug — shim tries to call canonical UMP subprocess without first running initialize, so remember/revise/forget/get/feedback all fail with "unknown tool". Out of scope here but flagged.