A skill memory layer that watches agent sessions, waits for a task to complete or fail, runs an LLM distillation pass, and writes the result as agent-skill-format Markdown files rather than embedded facts. Retrieval is tool-driven progressive disclosure (list_skills / get_skill / get_skill_file) reasoned about by the agent, not semantic similarity search. Skills export as a ZIP for reuse across frameworks. Backend is FastAPI over Postgres, S3, Redis, and RabbitMQ, reachable by REST from Python and TypeScript SDKs or a web dashboard.
- Storage
- PostgreSQL for metadata and state, S3-compatible object storage for the skill-file blobs, Redis and RabbitMQ for session and task queueing. The memories themselves are plain Markdown files following a SKILL.md schema — git-diffable and human-editable.
- Retrieval
- Tool-call-driven progressive disclosure: the agent calls get_skill or get_skill_file to pull whole files rather than chunks. There is deliberately no vector or embedding search path — the project's stated position is 'Git, grep, and mount to the sandbox' over semantic ranking.
- Self-host
- Self-host: heavy
- License
- Apache-2.0
- Pricing
- Self-hosted core is free and Apache-2.0. A managed cloud exists at dash.acontext.io offering free credits; notably, a 2026-04-21 commit removed all billing and Stripe integration and the /pricing page (now 404), defaulting new organizations to a free plan — so the cloud product currently has no visible paid tier. · Freemium
- GitHub stars
- 3,664
- Last release
- 2026-04-08
- Last commit
- 2026-04-21
- First catalogued
- 2026-07-25
Strengths
- Novel memory representation — plain, git-diffable, human-editable Markdown skill files instead of an opaque vector store
- First-class Claude Code plugin with real lifecycle hooks, more turnkey than an MCP-only integration
- Portable by design: skills export as a ZIP and are explicitly framework-agnostic
- Backed by memodb-io, the same team behind Memobase, suggesting real engineering depth
Watch out
- Apparent dormancy: no commits on any of 19 branches and no package releases since 2026-04-21 — over three months of silence, immediately after the commit that stripped billing; carded as needs-review rather than active
- Pricing is in flux — the vendor pulled its own /pricing page and Stripe billing in that same last commit, so any pricing statement here may go stale fast
- Retrieval is narrower than most memory layers: whole-file tool fetch with no semantic search, so recall-quality benchmarks are not apples-to-apples against vector-based systems
- Self-hosting is heavier than the 'quick proof-of-concept' framing implies — Postgres plus S3 plus Redis plus RabbitMQ is a real multi-service stack
- Easy to conflate with Memobase from the same org; they are separate products with different retrieval philosophies (Memobase models user profiles, Acontext stores skill files)
Best for
- Coding agents that should learn from mistakes and reuse debugged procedures as inspectable files
- Teams wanting memory that is auditable and git-diffable rather than an opaque embedding store
- Multi-framework shops wanting portable, no-lock-in skill files
How it integrates
Benchmark results
No sourced results yet.
Sources
- Monthly card re-verification — GitHub API repo metadata: 3,598 stars, license Apache-2.0 confirmed, archived=false, last commit on the default branch 2026-04-21 (third-party)
- Repo metadata — 3,585 stars, Apache-2.0; no commit on any branch after 2026-04-21 despite a later pushed_at (tag push, not code) (vendor)
- README — skill-memory model, distillation pipeline, Apache-2.0 license section (vendor)
- What is Skill Memory — confirms progressive-disclosure retrieval with no embedding search (vendor)
- Claude Code plugin README — stdio MCP server, five tools, skill sync into ~/.claude/skills/ (vendor)
- Python SDK on PyPI — v0.1.22, no release since 2026-04-08 (vendor)
- Memobase repo — confirms Acontext is a distinct product from the same org, not a rename or subproject (vendor)
Last verified 2026-08-01 · updated by refresh-framework-cards