Nahuali

Memory an agent can inspect before it trusts.

Memory retrieval usually begins with what context looks relevant? Nahuali also exposes four separate questions:

• What observation supports this memory?

• What conflicts with it or makes it stale?

• Do the current recorded-history checks pass?

• Do the available evidence and health checks support using this result?

The result is a local-first memory engine with deterministic trust verdicts, tamper-evident recorded history, and portable evidence. Its core does not need a model, account, API key, hosted service, or Docker.

A 60-second tour

Install the macOS or Linux binary. The installer requires the matching SHA-256 asset; the pinned path below can additionally require its Sigstore bundle.

curl -fsSLo /tmp/nahuali-install.sh \
  https://raw.githubusercontent.com/Arakiss/nahuali/main/scripts/install.sh
sh /tmp/nahuali-install.sh
export PATH="$HOME/.nahuali/bin:$PATH"

Run nahuali --version in a terminal to see the real axolotl spritesheet come alive beside the build identity. When the command is captured or redirected, it emits the stable nahuali <semver> line expected by agents, install checks, and benchmark tooling.

Record an observation, derive a claim from it, and require evidence at recall:

nahuali remember "Lena owns the release notes" --mention Lena --tag product
nahuali claim Lena owns "release notes" --source-last --confidence 0.92
nahuali recall "Who owns the release notes?" --authority --require-evidence
nahuali explore

nahuali explore is not just a record browser. It keeps three independent questions visible:

Press / to search every displayed memory-item field, Tab to filter by memory kind, and j/k to inspect the evidence behind a result. The axolotl watches over the memory and changes with its status. Rebuild the GIF above from real Ghostty windows on macOS with scripts/render-readme-tui-gif.sh; the canonical README asset is never replaced by the lower-fidelity text fallback.

Run nahuali demo for a narrated, non-mutating explanation of how Nahuali detects a specific in-place record change and keeps unsupported memory from driving action. The demo states which changes require an externally retained reference rather than relying on the live store alone.

Related MCP server: memotrust

Where Nahuali fits

Choose Nahuali when an agent needs local or operator-controlled memory and the decision to use a result must carry evidence, health signals, and explicit integrity limits. It is especially useful as a reliability layer around consequential memory rather than as a replacement for every retrieval system.

Choose an established recall-first or managed memory platform when the main requirement is hosted operation, a broad framework ecosystem, or publicly disclosed LoCoMo/LongMemEval answer-quality results. Nahuali does not claim leadership on those axes today. An application can also pair another retrieval system with Nahuali's evidence and governance checks.

What Nahuali checks

Evidence is part of memory

Observations are first-class episodes. Claims, relationships, procedures, and intentions can point back to the episode that supports them. Recall can refuse results without that path instead of filling the gap with confidence language.

Verdicts stay separate from relevance

Authority-aware recall carries one of four deterministic verdicts:

A certify verdict proves neither truth nor authorship. It means the available evidence and content-health checks passed; inspect HISTORY and EXTERNAL separately for internal history checks and an optional external comparison. Nahuali keeps those limits machine-readable instead of hiding them in documentation.

History is inspectable

The authoritative ledger is append-only and hash chained. Merkle inclusion and consistency proofs make verification compact; Ed25519 checkpoints bind a tree size, root, chain tip, origin, and ledger lineage to an external operator policy.

Nahuali is not a blockchain. It applies the transparency-log primitives standardized in RFC 9162 without publishing private memory to a public network or requiring consensus. Independent witness co-signing is a future step, not a capability claimed today.

One claim can travel with a receipt

Portable claim receipts contain only the selected claim, its evidence episode, an optional source envelope, their Merkle inclusion paths, and one signed checkpoint. Verification is offline and requires a policy held separately from the receipt:

CLAIM_ID="$(nahuali data --json | jq -r '.claims[-1].id')"
nahuali receipt-export \
  --claim-id "$CLAIM_ID" \
  --checkpoint checkpoint.json \
  --policy policy.json \
  --output claim-receipt.json

nahuali receipt-verify claim-receipt.json --policy policy.json

The verifier reports receipt_integrity separately from content_authority. Ledger commitment never becomes a claim that the remembered statement is true, that its author is authentic, or that an external source still contains the same bytes.

Offline receipt verification checks only the selected envelopes, their Merkle paths, provenance links, and checkpoint authorization. It trusts the authorized signers' commitment to that root; it does not replay or validate the complete ledger prefix. Use checkpoint-verify with the ledger for that stronger check. Receipts contain the selected memory verbatim and should be handled as sensitive data, not published by default.

Built for agents, operable by humans

The CLI is canonical. The TUI, stdio MCP server, local HTTP API, and Rust crate all use the same deterministic engine:

Prefer the CLI for terminal-based agent workloads. It is the direct, scriptable interface, supports batch commands without an MCP tool-call round trip, and makes the exact database and JSON contract explicit in every invocation. MCP is an adapter for clients whose integration boundary is a tool protocol. Both interfaces have the same trust semantics; choosing the CLI does not bypass validation. An embedded database has one process owner, so do not point a running MCP server and CLI processes at the same embedded database simultaneously. Use remote SurrealDB when independent processes need shared access.

Run nahuali init to install the bundled agent skill where supported and print a native MCP configuration. Nahuali is also published as io.github.Arakiss/nahuali in the official MCP Registry. See MCP onboarding for native and container configurations.

The HTTP API is unauthenticated by design and must not be exposed to an untrusted network.

Storage and recovery

memory_record in SurrealDB is the source of truth. The current-memory view, graph tables, snapshots, and semantic vectors are derived and rebuildable.

• A normal write appends one ledger event, updates the in-memory view, compares the desired graph with the last verified graph projection, and writes only rows that are new or changed. Rows no longer present are removed.

• If the stored projection checkpoint, schema version, error state, or content manifest cannot be trusted, Nahuali clears the derived graph and rebuilds it from the ledger. Recovery never rewrites authoritative history.

• Embedded SurrealKV is the zero-service default.

• Remote SurrealDB supports deliberately shared deployments.

• Lexical recall works without Qdrant or an embedding model.

• Qdrant and local embeddings are optional semantic tiers.

• Graph projection v2 fences concurrent rebuilds and validates a canonical content manifest for every ledger-derived projected table, not only row counts. SurrealDB projection-backed entity, timeline, pending-work, and health reads fail closed while a rebuild is active or when the manifest, schema version, or ledger tip drifts.

• Semantic validators detect vector-index drift before that derived tier is trusted.

• Backup restore replays the ledger and rebuilds derived state.

The embedded store has one process owner. A second process fails clearly rather than waiting indefinitely or risking concurrent writes.

Evidence and evaluation limits

The adapter-based Agent Memory Trust Benchmark is designed for cross-product evaluation of provenance, abstention, contradiction, staleness, non-mutating inspection, and tamper detection as separate cases. The checked-in result is first-party Nahuali evidence, not an independent comparison. The retrieval benchmark publishes every ranked item and latency sample for a versioned 24-memory, 12-query corpus.

The checked-in 0.8.0-beta.9 results are first-party, version-matched source builds. They are bound to a binary SHA-256 and source revision, but they are not claimed to be the published release archives. New published-release results must additionally match the release tag, target, archive name, archive digest, and exact binary through scripts/verify-benchmark-artifact-identity.py.

The small retrieval corpus is a regression gate, not a state-of-the-art claim and not a substitute for LoCoMo or LongMemEval. The broader checked-in governance suite exercises trust-specific failure classes. Run the controlled-beta gate with:

bash scripts/verify-controlled-beta.sh

Maintainers can verify the required GitHub repository settings separately:

NAHUALI_VERIFY_GITHUB_SETTINGS=1 bash scripts/security-supply-chain-check.sh

Beta boundaries

• Evidence proves traceability, not factual truth.

• A trusted checkpoint proves a committed ledger state, not an independent time.

• Detecting rollback or a fully re-chained history requires retaining an external checkpoint and policy.

• Scope labels organize contexts; they are not access-control boundaries.

• Self-inspection proposes work but never rewrites memory autonomously.

• Accounts, hosted sync, billing, and a managed control plane are not included.

• APIs and storage behavior may still change before 1.0.

Read BETA.md before using irreplaceable data and the full trust model before treating a verdict as an authorization boundary.

Install a pinned release

The short installer above always requires the matching SHA-256 asset. For a version-pinned install that also requires the Sigstore bundle, choose a release tag and run the installer from that same tag:

VERSION=vX.Y.Z-beta.N
curl -fsSLo /tmp/nahuali-install.sh \
  "https://raw.githubusercontent.com/Arakiss/nahuali/${VERSION}/scripts/install.sh"
NAHUALI_VERSION="$VERSION" NAHUALI_REQUIRE_SIGSTORE=1 \
  sh /tmp/nahuali-install.sh

This path requires cosign and stops if the archive checksum, signing identity, or bundle cannot be verified. See release verification for the exact identity, provenance, SBOM, and manual verification path.

Build from source

cargo build --workspace
cargo test --workspace
cargo install --path crates/nahuali-cli --locked

Docker is needed only for the optional remote development stack and Qdrant.

Go deeper

• Trust model

• Release verification

• Self-repair contract

• Governance benchmarks

• Security policy

• Contributing

Questions and design feedback belong in GitHub Discussions. Bugs and benchmark contributions have structured issue templates.

License

Nahuali is source-available under FSL-1.1-MIT, the Functional Source License 1.1 with an MIT future grant. You may inspect, use, copy, modify, self-host, and redistribute it for permitted purposes. Offering the current code as a competing commercial product or service is restricted; each published version converts to MIT after two years. The license does not prevent an independent reimplementation of the product's ideas. Read the practical licensing FAQ; the license text remains the binding source.