Active work
Resume from an objective, verified progress, remaining work, a blocker, and an exact next action.
Agent Needle
Repository-backed agent memory
Persistent context.
Surgical execution.
Give coding agents a durable thread through every session—plain Markdown memory, explicit boundaries, and the discipline to make the smallest verified change.
Repository
.
├── AGENTS.md
├── CLAUDE.md
├── .agent/
│ ├── IDENTITY.md
│ └── CONFIG.md
└── .memory/
├── ACTIVE.md
└── INDEX.mdHow it works
Memory lifecycle
Context that earns its place.
Agent Needle retrieves narrowly, checks the present, and preserves only context that improves the next decision.
Explore the memory modelStep 01RetrieveRelevant contextStep 02VerifyCurrent realityStep 03ExecuteSmallest changeStep 04CheckpointDurable contextOnly verified context moves forward.
repository memoryMemory with a defined purpose
Canonical modelDecisions
Preserve a consequential choice with its context, alternatives, rationale, consequences, date, and status.
Verified state
Record current facts about architecture, commands, integrations, constraints, and conventions.
Episodes
Keep dated, append-only records of meaningful actions, checks, failures, decisions, and unresolved threads.
Canonical session flow
The repository carries the thread.
Every non-trivial session starts by loading the defined context, then checking remembered state against the current repository before acting.
BootstrapRetrieve narrowly
Read identity, configuration, active work, and only the indexed memory relevant to the request.
Then inspect the actual repository.CheckpointPreserve verified context
Record what changed, what was checked, failures, remaining work, and an exact resumption point when work is unfinished.
Persist only with current authorization.Compatibility
One convention. Multiple agents.
Use the canonical `AGENTS.md` directly with Codex and compatible tools. Add the supplied `CLAUDE.md` adapter for Claude Code—without duplicating policy.
Start with a thread