An in-process, local-first Python memory library (`pip install archon-memory-core`) whose thesis is that memory should get better the longer it is used. Built on ChromaDB + Ollama, it pairs ranked top-1 retrieval with supersede-aware nightly consolidation, type-aware salience, an entity graph, active forgetting, and full replay/observability. Positions itself as a memory policy library (not an agent runtime), with LangChain and LlamaIndex adapters.
- Storage
- ChromaDB vector store (upsert on every add), optional Hindsight retain with graceful fallback. A WorkingMemory holds 4–7 slots (Miller's Law) and flushes to long-term store. Nightly consolidation (local Mistral/Qwen via Ollama) clusters by source/type/entity co-occurrence, compresses clusters into semantic facts, resolves contradictions toward newer truth, and archives originals (soft delete). A ForgettingPolicy applies salience decay; typed chunks (e.g. `credential`, `lesson`) never decay while `session` decays aggressively.
- Retrieval
- A six-stage pipeline: cosine retrieval over a 4x candidate pool → salience + recency scoring (adaptive per query type) → optional cross-encoder re-ranking (ms-marco-MiniLM) → MMR diversity selection (λ=0.7) → atomic-fact augmentation → dynamic tail pruning. An entity graph adds 2-hop expansion; every retrieval is replayable back to source chunks.
- Self-host
- Self-host: moderate
- License
- Apache-2.0
- Pricing
- Open-source Apache-2.0 ('free forever'), self-hosted (ChromaDB + optional Ollama for consolidation). Paid observability/eval/hosted tiers are described as on the roadmap (not yet available). · Free / OSS
- GitHub stars
- 6
- Last release
- 2026-04-18
- Last commit
- 2026-04-22
- First catalogued
- 2026-06-28
Strengths
- Supersede-aware consolidation: contradicting facts are archived with a link to what replaced them rather than left to compete at retrieval time
- Type-aware salience (high-value facts like credentials never decay) plus active forgetting, an entity graph, and replayable retrieval for observability
- Local-first (ChromaDB + Ollama), Apache-2.0, in-process library with LangChain/LlamaIndex adapters and an included eval harness
Watch out
- Headline accuracy (99.2% top-1) is measured on AMB — the project's OWN preregistered benchmark — so it is strongly self-reported; record via harvest-benchmarks with selfReported:true, not here
- Very low adoption (~6 stars) and not recently updated (last commit 2026-04-22); newest tags are pre-release AMB-v2 previews
- GitHub classifies the license 'NOASSERTION', but the LICENSE file is verbatim Apache-2.0 (confirmed); consolidation requires a local Ollama model
Best for
- Agents that accumulate contradictory facts over long horizons and want a local, consolidating memory library with built-in forgetting
Benchmark results
No sourced results yet.
Sources
- archon-memory-core README (vendor)
- GitHub API repo metadata (stars, release/tags) (third-party)
- LICENSE (verbatim Apache-2.0 text; GitHub auto-classifies NOASSERTION) (vendor)
Last verified 2026-06-28 · updated by discover-frameworks