mcp server • rust • memory

octobrain

hybrid search • sleep consolidation • knowledge boxes

give your ai assistant a long-term memory that outlives the context window — 22 memory types, hybrid BM25 + vector search, and a sleep-consolidation pass that folds a finished goal into one summary instead of leaving a hundred loose notes behind. retrieval quality is a measured, reproducible BEIR benchmark, not a claim.

brew install muvon/tap/octobrain

GitHub

⭐ 20

learn more

BEIR-benchmarked· HyDE query expansion· Git-scoped· Apache 2.0

22 memory types

4 mcp tools

+2 nDCG@10 vs bare embedding

10-30% recall lift from HyDE

key features

💾

persistent semantic memory

llms don't actually remember — they re-read every message from scratch and drop context when the window fills. octobrain stores memories externally with vector embeddings and hybrid BM25 + vector RRF fusion, so your ai retains critical context across sessions, not just within them.

😴

sleep consolidation

close a goal and octobrain folds every contributing memory into one archived summary — sleep-consolidate auto-clusters similar recent memories by similarity threshold, the same way biological sleep consolidates the day into fewer, stronger memories instead of a thousand loose fragments.

📦

knowledge boxes

import a git-backed knowledge bundle — box import <repo>, scoped to one project or --global across all of them — and octobrain box sync pulls and re-indexes it. shareable, versioned context a whole team subscribes to, not memory trapped in one machine.

🎯

query expansion (HyDE-lite)

pseudo-relevance feedback rewrites a query before search runs, lifting recall 10-30% on the long-tail queries where exact wording rarely matches how a memory was originally phrased. on by default, alongside the cross-encoder reranking stage.

🔗

auto-linking & memory graph

memories connect to related memories automatically via semantic similarity, or explicitly (depends_on, related_to). memory graph walks the connections multi-hop. Ebbinghaus temporal decay fades stale memories while frequently accessed ones stay strong.

📚

knowledge base indexing

index urls, PDFs, and docs (.txt, .md, .pdf, .docx, .html) into chunked, searchable storage — knowledge search finds passages by meaning, knowledge match runs a regex pass, and a stale source auto-reindexes the next time it is searched.

why octobrain?

ai forgets everything between sessions.

✓

persistent vector storage means memories survive across conversations. context that matters is always available — not re-explained every time.

a finished project leaves behind a hundred loose, unsearchable notes.

✓

consolidate folds a goal's contributing memories into one archived summary; sleep-consolidate auto-clusters similar recent memories the same way.

team knowledge lives in one person's local memory store.

✓

knowledge boxes are git-backed and shareable — import once, box sync keeps every subscriber current, scoped per project or global.

"better retrieval" is usually a claim, not a number.

✓

octobrain ships a reproducible BEIR benchmark against its own real retrieval path — hybrid search beats classic BM25 and the bare embedding on both measured datasets, numbers anyone can rerun.

mcp tools

available via Model Context Protocol for ai assistants

memorize store a memory with metadata; related_to links it inline at creation

remember semantic search with filters, returns 1-hop graph neighbors

forget delete memories, with confirmation required

knowledge unified knowledge-base tool — search, store, delete, read, match via one command field

when to use octobrain

cross-session context

your ai assistant remembers decisions, patterns, and preferences from previous sessions. no more repeating "we use PostgreSQL, not MySQL" every conversation.

a finished project becomes one summary

consolidate a goal and every memory that contributed to it archives into a single record. six weeks of debugging notes become one paragraph — searchable, not scrolled through.

team knowledge shared, not re-explained

publish a knowledge box once — architecture decisions, runbooks, onboarding docs — and every teammate's ai subscribes to the same git-backed source with box sync.

research & documentation recall

index api docs, design specs, and competitor analysis. your ai references previously read material without re-fetching — instant recall of anything it has processed.