MemoryAtlas

Peer-centric memory built on four primitives: Workspaces (app-level isolation), Peers (any long-lived entity — user, agent, group, project, or idea), Sessions (many-to-many interaction threads between peers), and Messages (data units that trigger reasoning). Honcho 3.0 (Jan 2026) split reasoning into three stages run by the deriver worker process: ingestion reasoning extracts deductive (explicit) and inductive (inferred) conclusions in parallel on write; an async Dreaming Agent handles lower-priority background work (summarization, peer-card generation, surprisal-prioritized reprocessing) off the request path; a query-time Dialectic Agent answers chat() calls with an agentic search-and-synthesize loop across five reasoning depths (minimal to max). Configurable LLM backends (OpenAI, Anthropic, Gemini); queries target this derived conclusion/representation layer rather than raw message history.

Deep diveHoncho deep dive: how theory-of-mind user modeling actually works in productionRead →
Storage
Managed cloud at api.honcho.dev, or self-hosted FastAPI server. Persistent layer: PostgreSQL + pgvector (default vector backend); Turbopuffer and LanceDB supported as alternatives. Optional Redis caching for hot paths.
Retrieval
Hybrid BM25 + vector search scoped to workspace, session, or peer level. Context endpoint aggregates messages, conclusions, and session summaries within a caller-specified token budget. Representation endpoint delivers low-latency static snapshots of derived peer insights. Chat API provides reasoning-grounded natural-language responses against the live peer model.
Self-host
Self-host: heavy
License
AGPL-3.0
Pricing
$2/M tokens ingested (storage + background reasoning included); retrieval unlimited; Reasoning API $0.001–$0.50/query by depth tier; $100 free credits on signup; startup program: $1,000 credits + 12 months subsidized pricing; enterprise custom · Freemium
GitHub stars
6,557
Last release
Last commit
2026-08-06
First catalogued
2026-06-28

Strengths

  • Theory-of-mind user modeling: derives conclusions about beliefs, preferences, and contradictions rather than replaying raw chat history
  • Peer-centric multi-entity model: any entity (user, agent, group, project, idea) is a Peer; tracks what one peer knows about another across many-to-many sessions
  • Broad integration surface: remote MCP (Claude Code, Cursor, Windsurf, VS Code, Zed, Cline, Codex, Goose), Python + TypeScript SDKs, REST API, and a first-party Claude Code plugin
  • Hybrid retrieval: BM25 + vector search combined with a context endpoint that aggregates conclusions within caller-specified token budgets
  • Tiered reasoning cost: $2/M for ingestion; per-query Reasoning API ($0.001–$0.50) lets callers pay only for the inference depth they need

Watch out

  • Background reasoning fires LLM calls on ingestion (and dreaming/dialectic calls on top) even when self-hosted — adds latency and cost regardless of hosting mode; AGPL-3.0 restricts commercial embedding without a commercial license
  • Git tags exist (e.g. v3.0.12) but there are no published GitHub Releases or changelog notes — pinning a version means tracking tags yourself, not reading release notes
  • Heavy self-host footprint: PostgreSQL + pgvector + Redis + deriver worker process + an LLM backend all required; multi-service orchestration is non-trivial for small teams
  • No first-party endpoint to wipe a peer's accumulated memory (a 'peer delete' endpoint is on the roadmap as of Apr 2026); the documented workaround is versioning peer names or deleting directly in Postgres, and headline benchmark scores (LongMemEval 90.4, LoCoMo 89.9) are self-reported without a disclosed embedder

Best for

  • Personalization that must model a user's evolving beliefs, preferences, and contradictions over many sessions (theory-of-mind, not fact recall)
  • Multi-agent or multi-party systems where what peer A knows about peer B — or what an AI agent knows about a human — needs to persist and update
  • Coding-assistant memory (Claude Code plugin, Cursor MCP) where context must survive across sessions, directories, and projects

How it integrates

Benchmark results

BenchmarkValueBackboneTrustSource
longmemeval90.4 accuracySelf-reportedHoncho (Plastic Labs)
locomo89.9 accuracySelf-reportedHoncho (Plastic Labs)

Sources

Last verified 2026-08-01 · updated by refresh-framework-cards