memory-bank

Memory Bank

Conversations → Knowledge Graph. Claude Code conversations become searchable, structured knowledge.

What's New in v1.4.x

• 🧾 Injection pipeline v2 — session dedup ledger (v1.4.0) — a fact is now injected at most once per session. Before: 74% inject rate × 5.5–8 facts × ~140 chars ≈ ~470 tokens per prompt, with the same facts re-injected across a session (~10k tokens per 30-prompt session). A bounded per-session ledger (400 ids, 7-day TTL, atomic writes, fail-open) filters already-injected facts; a repeated topic's 2nd occurrence injects 0 bytes (status:"deduped")
• ✂️ Token budget — per-fact 160-char truncation + 1,000-char block budget (lowest-relevance facts dropped first), plus an elapsed-budget gate on repeated-prompt detection: its 313k-exchange vector search (p95 498ms) is synchronous and cannot be preempted mid-flight, so it is skipped entirely when injection has already spent >700ms
• 📈 Savings are measured, not assumed — the inject log gains chars (block size) and deduped (facts saved) fields, so real-world token savings accumulate in observable JSONL
• 🌌 Knowledge Galaxy (ui/relations/) — Three.js 3D visualization of the ontology: 32 domains / 4.2k categories / 24.7k facts / 27.8k typed relations as a navigable galaxy — fact search, per-type edge toggles (SUPPORTS/INFLUENCES/SUPERSEDES/CONTRADICTS), relation-navigating detail panel, adaptive performance (compositor-only labels, point-size cap, eco-mode DPR). node ui/relations/generate-data.mjs then serve statically — the data file stays local (personal facts, gitignored)
• 🩹 Dependency self-heal (v1.4.2) — claude plugin update non-deterministically skips npm install for the new cache dir, and cc-sync ships plugin caches to other machines without node_modules. The dep-free injection thin client now detects ERR_MODULE_NOT_FOUND and spawns a one-shot detached npm install (atomic marker prevents loops), so a broken install heals itself on the first prompt
• 📦 Packaging fix (v1.4.1) — v1.4.0 shipped a stale committed dist/; the release checklist now hard-checks git status dist/ is clean before tagging
• Earlier releases: see CHANGELOG — v1.3.x added ontology backfill batching (~25×), headless LLM spawn isolation, the classification attempt ledger, vec-index self-heal, worker flood caps, and the warm inject daemon (~2.3s → ~0.07s per prompt)

Features

• Knowledge Graph -- Ontology classification (Domain → Category) + typed relations (INFLUENCES, SUPPORTS, SUPERSEDES, CONTRADICTS)
• RAG Search -- Search results auto-enriched with related facts and ontology context
• Conversation Search -- Semantic vector search + FTS5 (BM25) full-text search across all past conversations
• Full-History Analysis -- memory-bank analyze + analyzing-all-conversations skill: coverage-checked report over the entire conversation index (projects, facts, domains, timeline, backfill gaps) — run memory-bank sync first so new conversations are indexed
• Context Injection -- Related past decisions are automatically prepended to every prompt (UserPromptSubmit hook): baseline-margin relevance gate + 1-hop ontology expansion + per-session dedup ledger and token budget (v1.4.0), with fail-loud JSONL observability logs
• Fact Extraction -- Automatic extraction of decisions, preferences, patterns from conversations (trivial-exchange filtering, in-session dedup, confidence gating, LLM call budgeting)
• Fact Consolidation -- Duplicate detection, contradiction handling, evolution tracking
• Graph Traversal -- Multi-hop exploration (up to 3 hops) to trace decision chains
• Cross-Project Insights -- Find similar decisions from other projects
• Fact Provenance -- Trace any fact back to its source conversation
• Scope Isolation -- Project facts stay in their project, global facts are shared
• Compressed Archive Support -- Transparent .jsonl.zst reads across every path (parser, read tool, search, sync, stats, verify) using Node's built-in zstd (Node >= 22.15) — archives compressed out-of-band keep working
• MCP Integration -- 9 tools: search, read, search_facts, search_ontology, ask_avatar, trace_fact, explore_graph, cross_project_insights, graph_stats
• 3D Visualization -- Interactive neon-style knowledge graph with data flow animation
• Web UI -- Dark-theme web interface for browsing and searching conversations

How It Works

Data Pipeline