CONTEXT AUTOPILOT · v0.1 · FREE & OPEN SOURCE

You already taught your agent everything.
It just wasn't listening.

Context Autopilot mines your real coding-agent sessions — every instruction you repeated, every correction you made, every action you rejected — and distills them into CLAUDE.md / AGENTS.md rules you approve. Automated context collection, grounded in evidence.

$ npx context-autopilot scan
Works with Claude Code & Cursor · Project + global rules · Zero dependencies · 100% local · No API key needed

How it works

Your session history is the highest-signal context source that exists — it's a literal record of what your agent got wrong and what you had to say to fix it. Autopilot turns it into durable context.

01 · OBSERVE

Mine your sessions

ctxlayer scan reads your local Claude Code transcripts and Cursor sessions and finds three kinds of evidence: instructions you repeated across sessions, corrections you made after the agent went wrong, and tool calls you rejected.

02 · DISTILL

Evidence in, rules out

ctxlayer distill sends the signals — not your whole history — through Claude and gets back imperative, project-specific rules. Each one carries the quotes that justify it. Generic advice gets filtered out.

03 · APPROVE

You stay in charge

ctxlayer apply shows each proposal with its rationale and evidence. Accepted rules land in a managed block inside CLAUDE.md and AGENTS.md — your hand-written content is never touched.

Real output, real project

Run against an actual SaaS project's session history, Autopilot surfaced rules like these — each one a convention the developer had already taught the agent the hard way:

$ ctxlayer distill
Distilling 26 signal(s) with your local `claude` CLI…

[1/8] Perform click-and-type tests before reporting any UI work complete (confidence: high)
  + Before declaring any screen done, click every button and verify it works —
    do not ship cosmetic (non-functional) buttons.
  evidence: "There are still so many buttons that dont work, like the publish…"

[2/8] Staff login cannot access or toggle to admin view (confidence: high)
  + When authenticated as staff, the admin role toggle must be hidden —
    this is access control, not a UI preference.
  evidence: "…from the staff login, I do not want to see the admin view."

[3/8] The platform ships as a self-hosted Docker image (confidence: high)
  + Clients run the image on their own server; never store or pull their
    data into any external cloud platform.
  evidence: "the data for their schedules and rosters should stay with them…"

Next: `ctxlayer apply` to review and write the ones you accept.

Why evidence-based, not generated

Auto-generated context files make agents worse

Research on LLM-generated CLAUDE.md/AGENTS.md files found they reduce task success and raise inference cost — repo scans produce generic filler the model then has to wade through. A focused 50-line file beats a sprawling 1,000-line one.

Your corrections are ground truth

When you tell an agent "no — like this" you are labeling training data for free. Autopilot only proposes rules your own words support, quotes them as evidence, and prefers fewer, higher-confidence rules over volume.

Works with every agent — and stays fresh

Output lands in both CLAUDE.md and AGENTS.md, so Claude Code, Cursor, Copilot, Codex, and 30+ other agents all benefit. And ctxlayer stale catches the reverse problem: context your repo has outgrown — missing files, removed npm scripts — before it misleads an agent. CI-friendly (exit 1 on findings).

It learns you, not just your repo

ctxlayer distill --global mines all your projects across all your tools for rules about how you work — "explain in plain English", "don't build while I'm brainstorming", "parallelize independent tasks" — and maintains them in your personal ~/.claude/CLAUDE.md. No other tool does evidence-based personal context. It's the first step toward the endgame: agents that absorb how you work without you ever "building an agent."

The Context Layer Index

An independent, curated map of the context layer — who does what, without the vendor slant. Updated continuously.

Agent memory layers

Mem0Managed long-term memory; largest ecosystem
ZepTemporal knowledge graph; strong on time-aware recall
LettaOS-style paged memory (MemGPT lineage)
SupermemoryPersonal memory vault with deep MCP integration
CogneeGraph + vector memory pipelines
GraphitiOpen-source temporal graph engine
LangMemMemory inside the LangGraph loop

Context for coding agents

Context AutopilotEvidence-based context distilled from your sessions — this site
AGENTS.mdThe open cross-agent context file standard
Context7Up-to-date library docs as agent context
GitHub MCPRepo, PR & issue context over MCP
Claude CodeCLAUDE.md, rules, skills, hooks — deepest native support

Enterprise context platforms

GleanFully-managed enterprise context & search
AtlanGoverned data-catalog context layer
InterloomCaptures expert operational knowledge for agents
GraphlitManaged ingestion → agent-ready knowledge
LlamaIndexRetrieval framework powering many context stacks

Observation & capture

Screenpipe24/7 local screen+audio capture (source-available)
PiecesDeveloper workstream memory across IDE/browser
LimitlessWearable + meeting capture (acquired Rewind)
Codex Record & ReplayDemonstrate a task once → reusable skill
MindedRecord a browser task → team workflow
ScribeAuto-generated SOPs from screen recordings

Context observability

LangfuseOpen-source LLM traces, evals, prompt mgmt
BraintrustAgent observability + evals, IDE-native via MCP
LangSmithTracing & debugging for LangChain stacks
Arize PhoenixSelf-hostable, OpenTelemetry-native traces
mcpsnoop"Wireshark for MCP" — see what agents actually receive

Memory portability

MemoryXExport/import memories across ChatGPT, Claude, Gemini
AI Context FlowUniversal memory layer above every assistant
MCPThe protocol the whole layer speaks

The Context Layer, weekly

One email a week on who's winning the context layer — new tools, benchmarks, and what actually works. No filler.