LNM Brain MCP Server

Search Second Brain. Returns entity facts FIRST (instructions at top), then doc matches. Best for recall. Semantic/vector fusion + 1-hop run BY DEFAULT (bridges synonyms, e.g. supplements≈medications). Optional category filter: decision|bugfix|feature|discovery|conversation|note. Set fast=true ONLY …

Get confidence-scored facts for a specific entity. Supports category filter: fact|preference|instruction|history. Resolves aliases automatically.

Natural language Q&A over your entire Second Brain. Uses semantic search + reranking + entity facts + instructions. Best for questions.

Semantic vector search using AI embeddings.

Save content. Auto-extracts entity facts (with categories + triples) + contradiction detection + normalization. surface is FIRST-CLASS — pass it on every capture. If YOU (the assistant) can compress, pass wiki_body with a clean wiki-formatted summary that preserves all nuance — the worker stores it …

Full Karpathy INGEST pipeline + fact extraction + triple storage. Pass wiki_body to store an assistant-compressed summary verbatim and skip the weak external-LLM pass. type='project' routes to wiki/projects/ with project_meta frontmatter.

Return the freshest SESSION-HANDOFF entry as {observation_id,timestamp,timestamp_ist,surface,topic,state,next_action,trail}. KV-only, no LLM. Call this FIRST at session start.

Recent observations sorted by timestamp DESC across ALL session_id buckets. Optional filters: surface, since (ISO/date), until (ISO/date), entity. KV+graph join, no LLM.

History of SESSION-HANDOFFs sorted by timestamp DESC. Includes superseded ones. Useful for week/month reviews. KV-only, no LLM.

Pre-computed snapshot of the most recent observations across all surfaces (refreshed by cron). Zero KV reads beyond one. KV-only, no LLM.

Health-check: orphans, uncompressed files, entity facts summary, triple count.

Read full content of a specific file.

Get knowledge graph nodes + edges.

Write a file at a specific path.

Query structured triples (subject|predicate|object) for an entity. Filter by predicate, category, or derivation. Direct lookup, no LLM needed. Each triple includes derivation ('stated'=verbatim from user | 'inferred'=LLM-extracted | 'reinforced'=re-confirmed | 'unknown'=legacy) and source_obs_id (or…

God-node ranking: the most-connected entities in the whole Second Brain right now, scored by fact_count + 2×triple_count. Surfaces what everything flows through — the hubs whose removal would fragment the knowledge graph. Like Graphify's god-nodes, but for personal cross-session memory. KV scan, no …

Shortest connection between two entities via the triple graph (BFS, ≤4 hops). Answers 'what connects X to Y across everything I've captured' with the hop-by-hop path + the predicate on each edge. Like Graphify's `path` command, but over personal knowledge. Returns null if unreachable. KV-only, no LL…

Lightweight index of observations in a session. Defaults to most recent session. ~40-50 tokens per entry. No LLM, KV-only.

Fetch full content of one observation by id (returned from get_session_index). On-demand, no LLM.

Synthesized self-model of Ritik: identity, beliefs, active state, and an adversarial Tensions/Open-Loops section that surfaces where recent actions contradict committed rules. Pre-computed by cron (daily 4am IST), augmented live with current session context. NOT a flattery doc.

Tiered personalized self-model of Ritik — TOKEN-CHEAP by default. depth:'core' (default) = ~10 identity facts + top behavioral + 1-line voice (tiny, load every session). depth:'domain' + domain:'writing|design|build|client|decision|habit' = one targeted slice (e.g. his ShreeTools ad-writing style) —…

Force-regenerate the About-Me self-model now (instead of waiting for the daily cron). Returns counts per layer + tensions found.

Get domain routing index — shows which content domains exist (code/work/personal/research/health/finance/general), their sizes, and recent titles. Optional surface filter narrows by capture origin. Use this FIRST to narrow search scope before querying.

v9: D1 FTS5 keyword search over observation title+content+entities+tags. Instant retrieval, no Vectorize lag. Falls back to GitHub keyword scan if D1 binding absent.

v9: Hybrid recall — runs semantic + keyword (FTS5) + entity + triple lookups in parallel, fuses via reciprocal rank fusion, applies recency boost. Returns top K with provenance (which methods matched).

v9: Orient at session start — single <3 KB payload with latest_handoff + open_threads + recent_verdicts + active_topics. Call this FIRST after BRAIN-STARTUP.

v10.2.2: Correct/retract wrong memory. mode='fact' retracts one entity fact (marks superseded, confidence→0, so it stops surfacing but stays auditable) — pass entity + fact (exact text). mode='triple' retracts one triple — pass entity + predicate (+ object to disambiguate). mode='observation' marks …