chimera-memory 0.20.0

Local-first reliability ledger for AI coding-agent work

chimera-memory

Local-first reliability ledger for AI coding-agent work.

Records what an agent tried, which command verified it, what happened, and what receipt proves it. Everything stays on your machine — no cloud, no sync, no account required.

Install

pip install chimera-memory

Requires Python 3.12+. Installs chimera-memory-types, pydantic, and filelock automatically.

Quickstart (5 commands)

# 1. Initialise the local ledger (also adds .chimera-memory/ to .gitignore)
chimera-memory init

# 2. Start a session
chimera-memory session start \
  --branch feat/my-branch \
  --task-label "fix type errors" \
  --agent kiro \
  --model claude-sonnet-4.6 \
  --harness-id kiro-cli

# 3. Wrap a verification command (test, lint, typecheck, …)
chimera-memory wrap \
  --failure-origin organic_real \
  --scope-path . \
  --verification-scope package \
  -- pytest tests/ -q

# 4. End the session
chimera-memory session end --status PASSED

# 5. View the receipt
chimera-memory receipt latest

Onboard an agent in 90 seconds

Get the full session/wrap/repair-loop protocol for your agent:

chimera-memory agent-guide --agent generic   # or: kiro, codex

Generate a copy-paste session scaffold for your package:

chimera-memory template dogfood --scope-path packages/chimera-memory

Prompt templates for sustained use are in docs/prompts/ in the source repo:

• kiro-dogfood.md — Kiro agent session discipline
• generic-agent-dogfood.md — Any agent
• release-closeout.md — Release closeout checklist

Health check

chimera-memory doctor
chimera-memory doctor --json

doctor is read-only. It checks:

• Is .chimera-memory/ initialized?
• Is .chimera-memory/ in .gitignore?
• Is there an active session?
• Claim count and integrity status
• M2B readiness level

Exit code: 0 = healthy, 1 = warnings (e.g. no active session), 2 = critical (e.g. not initialized).

Is my ledger healthy?

doctor also audits whether the ledger is being used correctly:

chimera-memory doctor        # shows Evidence Hygiene section
chimera-memory doctor --json # machine-readable, includes evidence_hygiene + next_actions

Evidence hygiene checks:

• Scoped claim ratio (claims with --scope-path set)
• Unknown failure_origin count (claims missing classification)
• Repair-loop completeness (loops with same_scope_after_fix)
• Orphaned repair phases (--repair-phase without --repair-loop-id)
• Test/synthetic and invocation-artifact claim counts

If issues are found, doctor prints actionable next steps.

Preflight Intelligence (v0.6)

Before starting work, surface historical failures, repair-loop lessons, and hygiene warnings:

chimera-memory preflight --from-git
chimera-memory preflight --scope-path packages/chimera-memory

Shows:

• Historical failures in the relevant scope (organic_real + controlled_real)
• Repair-loop lessons (what failed, was it fixed?)
• Hygiene warnings (invocation_artifact issues — not product defects)
• Recommended verification commands

Not M2B scoring or model ranking. Advisory only.

CI receipt bundle

Produce a full CI-ready artifact bundle (receipt, failures, reliability, preflight, GitHub summary):

chimera-memory receipt bundle \
  --output-dir ./ci-bundle \
  --include-preflight \
  --from-git

Upload ci-bundle/ as a CI artifact to preserve receipts across runs.

Evidence dry-run

Export claims as a portable evidence bundle and inspect what a merge would do:

chimera-memory evidence bundle --output-dir ./evidence
chimera-memory evidence import ./evidence --dry-run --json

Write-import is not available. Dry-run only.

M2B readiness

chimera-memory m2b-readiness

Shows whether the local ledger has accumulated sufficient repair-loop evidence to consider M2B (model-to-baseline) comparison. This is a local readiness gate, not a scoring system.

Run chimera-memory m2b-readiness --explain to see what evidence is missing before scoring can be trusted.

What is not built

Capability	Status
M2B drift scoring	Not built
Model ranking / routing	Not built
Hosted / cloud sync	Not built
Remote team ledger	Not built
Evidence write-import	Not built (dry-run only)
Dashboard or UI	Not built

Platform support

Platform	Status
macOS	✅ Tested
Linux	✅ Tested (CI and Docker)
Windows	⚠️ Not officially tested in 0.1.x

No intentional Windows lockout is known. If you encounter a Windows-specific issue, please open an issue.

Troubleshooting

preflight --from-git shows source: none

This is expected when the working tree is clean (no modified files). Use --scope-path <dir> to specify a target directory explicitly:

chimera-memory preflight --scope-path src/

receipt latest shows "No closed sessions"

You need to complete a session first:

chimera-memory session start --branch <branch> --task-label <label> --agent <agent> --model <model> --harness-id <id>
# ... run chimera-memory wrap commands ...
chimera-memory session end --status PASSED

m2b-readiness shows BLOCKED

The ledger needs at least 5 genuine organic_real failures followed by fixes to clear the gate. Run real verification cycles — do not manufacture failures.

verify shows LEGACY_UNSIGNED

Claims recorded before the integrity chain was introduced show as LEGACY_UNSIGNED. This is informational. Zero Broken records is the signal that matters.

Security note

chimera-memory wrap records the command string and bounded stdout/stderr. Built-in redaction removes tokens matching common secret patterns. However:

• Do not pass secrets, passwords, or API keys as literal command arguments.
• The .chimera-memory/ directory contains your local ledger. Do not commit it to version control.

Add to .gitignore (done automatically by chimera-memory init):

.chimera-memory/

Test fixture isolation

When writing tests that call chimera-memory wrap as a fixture, set CHIMERA_DQ_NO_WRITE=1 to prevent test runs from writing to the real local ledger:

CHIMERA_DQ_NO_WRITE=1 chimera-memory wrap --failure-origin organic_real -- pytest tests/

The wrapped command runs normally and the exit code is preserved. No claims are written. A notice is printed.

Do not set CHIMERA_DQ_NO_WRITE=1 for real dogfood verification sessions. It is for test fixtures only.

License

MIT — see LICENSE