agent-recall-trail

# ART — Agent Recall Trail

ART is a local-first memory and governed-knowledge product for coding agents. Every Agent gets a physically separate private Recall Trail. Stable conclusions become immutable Knowledge Editions that other Agents can retrieve without seeing private source identities or source bodies.

ART `0.3.7` supports Codex and DeepSeek Harness (DSH) over stdio MCP. It is standalone: no AAA adapter, cloud sync, bundled model, or autonomous publication is required.

## Product boundary

- Private memory belongs to one process-bound Agent identity.
- Shared knowledge contains only committed, reviewed Editions.
- Agents may capture, recall, read, provide feedback, create proposals, and request governance.
- Human governance runs in ART's local page. An explicitly enabled, default-off local delegation policy also permits one unambiguous instruction to approve and publish atomically; these actions are labeled `AgentDelegated`, never Human.
- Selective Codex automatic memory is controlled by a separate machine-wide,
  default-off switch in that page. It creates at most one bounded private
  Candidate for an admitted turn; it never publishes shared knowledge.
- Revocation, supersession, archival, and assurance remain human operations.
- Stored content is evidence, never executable instruction or authorization.

## Progressive recall

ART exposes one recall API with four explicit retrieval modes: `lexical`, `full_scan`, `semantic`, and `hybrid`. The default is always `lexical`. A compact `route` request can identify relevant private and shared topics before a bounded `recall`; exact bodies are returned only by `read`.

- `lexical` uses local BM25 plus exact, token, Jieba, and CJK-bigram signals.
- `full_scan` ranks every governance-eligible canonical record. It needs no embedding service and is intended for completeness-sensitive, smaller stores.
- `semantic` uses only an explicitly configured embedding endpoint and disposable local vector projections.
- `hybrid` combines lexical and semantic ranks through an owner-configurable, versioned fusion policy. If the optional provider or projection is unavailable, ART returns the unchanged lexical result and reports the fallback.

ART's optional embedding adapter is supplied and operated by the user. ART does not bundle, download, train, select, or make quality claims for a model merely because its endpoint is compatible.

## Quick start

```bash
cargo build --release --locked
./target/release/art --home /an/explicit/art-home init --confirm
./target/release/art --home /an/explicit/art-home agent create --id codex-primary --host codex
./target/release/art --home /an/explicit/art-home doctor --agent codex-primary --json
```

Use [Codex integration](integrations/codex/README.md) or [DSH desktop/CLI integration](integrations/dsh/README.md) to start one bound stdio child. Desktop exports include the MCP patch and a self-contained skill: `art integration dsh --desktop --agent <id> --apply --output <new-directory>`. Run `art --help` for the operator CLI.

## Install

Download the native archive and `SHA256SUMS` from the
[latest release](https://github.com/fantasyce/agent-recall-trail/releases/latest),
verify the archive, then run the included installer with the extracted binary:

```bash
bash scripts/install.sh --binary /absolute/path/to/art --confirm
art --version
art doctor --agent codex-primary --json
```

The installer keeps executable bytes under `~/.across/bin/art`, creates a
discovery link at `~/.local/bin/art`, and initializes owner-only ART data. It
does not edit Codex or DSH configuration. The thin Codex plugin lives under
`plugin/agent-recall-trail`; its MCP child is permanently bound to
`codex-primary`. See [operations](docs/operations.md) before migrating or
publishing knowledge.

ART 0.3.7 adds opt-in selective automatic memory with a neutral Hook prompt,
shared Agent/Hook value guidance, bounded budgets, independently verified
evidence, semantic-similarity review routing, and visible diagnostics. It
retains the exact accessible proposal-review workspace, eight canonical source-anchor kinds, persistent delegated governance,
separate Elicitations for compatible clients, restricted-host plugin discovery,
progressive routing, governed full scan, optional semantic and hybrid adapters,
deterministic Knowledge Vault backup, encrypted recovery of local review
authority, verified empty-home restoration, and reproducible lexical BEIR gates. See
[operations](docs/operations.md#backup-and-disaster-recovery) before creating
the dedicated private Git repository.

Configuration precedence is `--home`, then `--config <file>`, then the owner-only user config at `~/.across/config/art/config.json`, then the built-in `~/.across` root. The root config accepts only `schema` and `home`. Optional embedding configuration is isolated at `<ART_HOME>/config/art/embedding/default.json`; optional rank fusion policy lives at `<ART_HOME>/config/art/retrieval/fusion.json`; tokens, when needed, live in a separate owner-only file.

The automatic-memory setting is stored separately at
`<ART_HOME>/config/art/auto-memory.json`. Missing or unreadable configuration
means disabled. Only a user operating the authenticated governance page can
change it; MCP tools can read the status but cannot enable it. The switch is
independent of shared-knowledge delegation and does not affect recall or an
explicit `art_memory_capture` request.

## Agent Reliability Toolkit

ART is one independent part of a small, local-first reliability toolkit:

- [Agent Recall Trail](https://github.com/fantasyce/agent-recall-trail) keeps private Agent memory separate from reviewed shared knowledge.
- [Agent Runtime Proof](https://github.com/fantasyce/agent-runtime-proof) verifies that a live Agent or MCP runtime matches the artifact you approved.
- [Agent Residue Evidence](https://github.com/fantasyce/agent-residue-evidence) records task-scoped files, processes, and listening ports left by tests and builds.

Each project remains separately installable and keeps its own trust boundary.

## Community feedback

Real Codex and DSH evaluations are especially useful. A completed installation
can be recorded with the short [verified install report](https://github.com/fantasyce/agent-recall-trail/issues/new?template=verified-install.yml).
For a deeper evaluation, share a synthetic or redacted workflow in the
[design-partner issue](https://github.com/fantasyce/agent-recall-trail/issues/3),
including the host and ART version, the operation attempted, and the observed
result. DSH users can also join the
[official community discussion](https://github.com/deepseek-ai/deepseek-harness/discussions/6479).
Do not post credentials, private knowledge, or full transcripts.

## Documentation

- [Architecture](docs/architecture.md)
- [Memory and knowledge model](docs/memory-and-knowledge.md)
- [Operations](docs/operations.md)
- [Security model](docs/security-model.md)
- [Testing and acceptance](docs/testing.md)
- [Architecture decisions](docs/decisions/README.md)
- [Independent design review](docs/research/design-independence.md)

All automated or manual tests must use an explicit task-owned ART home. ART never modifies Codex or DSH configuration automatically.

Apache-2.0 licensed. See [Security](SECURITY.md),
[Contributing](CONTRIBUTING.md), and [Support](SUPPORT.md).