Local evidence
Raw, append-only, integrity-checked, and kept on your machine.
Unredacted notes, logs, traces, and artifacts stay local.
Evidence before memory
Memory that can show its work.
A local-first, evidence-backed continuous memory system for Codex, Claude Code, and OpenCode.
- ✓Preserve raw evidence locally in an append-only ledger
- ✓Promote only reviewed, redacted memories
- ✓Sync portable memory through private Git
- ✓Measure whether retrieval improves later tasks
Reviewed memory
Operator-attested, redacted, ready to share.
Policy requires human inspection; caller identity is not authenticated.
Private Git
Versioned memory under your control.
Portable, auditable, conflict-aware revisions.
Coding agents
Codex, Claude Code, and OpenCode.
Retrieve what helps. Record what was delivered.
Privacy boundary
Raw evidence never leaves your machine unless you explicitly create a separate encrypted backup. Git receives only promoted portable memory.
Evidence chain
Every promoted memory links back to local sources and review records. Provenance is preserved without publishing private transcripts.
Cross-agent retrieval
One verified memory format supports consistent retrieval across Codex, Claude Code, and OpenCode integration surfaces.
▥ Retrieval impact (frozen synthetic capability suite)
20tasks
19wins
1tie
0losses
Observed Codex results compare paired runs with and without verified memory across a sealed set of deterministic tasks. The aggregate receipt binds hashes and summary counts; raw execution evidence stays local and cannot be independently reconstructed from the public receipt.
Inspect the aggregate receipt →△ LimitationsThis capability suite is not longitudinal efficacy certification. Results may not generalize to different tasks, future models, or long-term use; the system reports insufficient prospective evidence as
not_evaluable.Strict where trust matters. Practical where work happens.
The system separates capture, judgment, distribution, delivery, and measurement so one convenient shortcut cannot silently become trusted memory.
Append-only local evidence
Source bytes, normalized events, completeness gaps, and capture receipts stay in a hash-chained local store.
Evidence-bound review packets
A concise local packet shows the exact candidate, support, conflicts, and content hash without combining validation and promotion attestations.
Verified memory loadouts
Named, immutable sets bind exact active revisions, agent eligibility, scope, and budgets for reproducible task context.
Prospective longitudinal studies
Requests, regular-file workspace snapshots, and acceptance hashes are sealed before execution. Archive and tree hashes are verified around launch; reused reservations, failed or missing trials, missing arms, and short duration remain not evaluable.
One evidence path, seven explicit boundaries.
Raw threads are not portable memory, a retrieved memory is not proof of adoption, and a green test is not longitudinal evidence.
Agent trajectory
Locally exposed task events and source bytes.
Evidence ledger
Append-only records, gaps, and integrity checks.
Episodes
Reconstructed task attempts and compaction checkpoints.
Candidates
Evidence-backed claims awaiting judgment.
Promotion
Separate operator attestation, redaction, immutable revision.
Private Git
Portable revisions and explicit loadouts.
Delivery
Scoped retrieval, injection, adoption, outcome.
Keep the evidence private. Carry the useful memory forward.
MIT licensed, cross-platform, and designed to fail closed when a claim cannot be verified.