MemoryAtlas

A workspace runtime that compiles heterogeneous sources (chat logs, documents, code, images, audio, tool traces) into three durable Markdown layers — Index (INDEX.md), Skill (SKILL.md), and Memory (MEMORY.md) — via a memorize() pipeline (ingest → preprocess → extract → organize → persist) and serves scoped, ranked layers back via retrieve(). The SKILL.md layer auto-extracts reusable workflow patterns from agent tool traces on every memorize() call, letting agents improve at recurring tasks without manual tagging.

Storage
Typed MemoryItems (profile/event/knowledge/behavior/skill/tool) grouped into auto-generated MemoryCategory folders with topic-level summaries, each item tracing back to the raw Resource it came from. Embeddings are indexed over folders, files, and sources; records persist through a configurable backend (inmemory, SQLite, or Postgres with pgvector).
Retrieval
retrieve() routes and scopes a query to relevant category folders, ranks matching items by relevance, and traces results back to source artifacts. Supports two strategies: 'rag' (vector-first with optional LLM routing and sufficiency checks) and 'llm' (LLM-ranked at each tier). The embedding index supports low-latency recall over the compiled workspace.
Self-host
Self-host: moderate
License
Apache-2.0
Pricing
Open-source Python library (memu-py, Apache-2.0), free to self-host with inmemory, SQLite, or Postgres backends. Cloud-hosted API at memu.so: free Starter tier (30 memory calls/period) and a paid Professional tier (600 calls/period). The public pricing page has gone: memu.pro/pricing and memu.so/pricing both returned HTTP 404 on 2026-08-01, and neither homepage lists tiers, so the tier limits above are carried over from the June 2026 capture and could not be re-confirmed — contact info@nevamind.ai for current limits and enterprise options. · Free + paid
GitHub stars
14,276
Last release
2026-03-23
Last commit
2026-08-09
First catalogued
2026-06-28

Strengths

  • Multimodal ingestion — text, documents/URLs, images/video (captioned), and audio (transcribed) are all compiled into one workspace
  • Markdown layers are inspectable and editable; every memory item traces back to the source artifact it came from
  • Self-evolving SKILL.md layer auto-extracts reusable workflow patterns from tool traces, improving agent behavior over repeated runs without manual tagging
  • Active project with a large community (~14k stars) and official SDKs for Python, JavaScript, Go, and Java
  • Pluggable storage backends (inmemory, SQLite, Postgres/pgvector) and configurable LLM provider routing for embeddings, vision, and transcription

Watch out

  • Cloud free Starter tier is capped at 30 memory calls per period — verify current limits at memu.pro/pricing before committing to the hosted API
  • License: repo LICENSE.txt is verbatim Apache-2.0 text but GitHub auto-classifies it as 'NOASSERTION' due to the non-standard filename — confirm there is no added rider
  • memorize()/retrieve() quality and benchmark numbers (92.09% on Locomo) are vendor-reported, not independently evaluated
  • Latest tagged release is v1.5.1 (2026-03-23); repo has been actively committing since but no newer release tag

Best for

  • Agents needing a multimodal, file-system-shaped memory workspace with source-traceable Markdown layers
  • Teams wanting a managed cloud memory API with a free starter tier before committing to self-hosted infrastructure
  • LangGraph or n8n workflows where a drop-in memory adapter with built-in skill extraction is needed

How it integrates

Benchmark results

No sourced results yet.

Sources

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