rta-smriti-brain

# Rta-Smriti Brain **A sovereign, local project-memory and evidence layer for AI coding agents.** Rta-Smriti gives every project a durable memory: repository structure, agent sessions, decisions, evidence, and the exact state needed to continue work without retelling the story.  [](https://github.com/sulabhdubey/rta-smriti-brain/actions/workflows/ci.yml) [](https://github.com/sulabhdubey/rta-smriti-brain/actions/workflows/binaries.yml) [](https://github.com/sulabhdubey/rta-smriti-brain/releases) [](LICENSE) [**Install**](#ten-minute-start) | [**Current release**](https://github.com/sulabhdubey/rta-smriti-brain/releases/tag/v1.1.0-alpha.6) | [**Live website**](https://sulabhdubey.github.io/rta-smriti-brain/) | [**Documentation**](#documentation) | [**Discussions**](https://github.com/sulabhdubey/rta-smriti-brain/discussions) > **Independent coverage:** [Harness Institute research: Rta-Smriti keeps agent memory local](https://www.harness.institute/research/rta-smriti-brain-keeps-agent-memory-local) | [Featured on The Next New Thing](https://www.youtube.com/watch?v=AWzzmrCPe-A&t=1350s) ## v1.1B Governed Federation **Current release: v1.1.0-alpha.6.** Safer imports and backups, clearer graph states, and updated build dependencies. Guided setup and read-only recovery remain the starting point, with optional encrypted collaboration over selected memory. Start with the [isolated Windows pilot guide](docs/PILOT_GUIDE.md). > **Current maturity:** `v1.1.0-alpha.6` is an advanced early-adopter release for Windows, > macOS, and Linux. It is useful for real projects, but it is not yet presented as a > broadly supported production platform. Read the bounded > [release verification record](docs/RELEASE_VERIFICATION.md). ## The Problem Every new agent session begins with an expensive question: **what is true about this project right now?** Repositories hold code, chats hold decisions, tools hold test results, and people hold the constraints. Ordinary retrieval collapses them into text. Rta-Smriti keeps them local, connected, time-aware, and evidence-labelled. ```mermaid flowchart LR A[Repository] --> B[Rta-Smriti Brain] C[Agent sessions] --> B D[Decisions and checkpoints] --> B E[Local evidence] --> B B --> F[Project Reality] B --> G[Governed context pack] B --> H[Fresh-session continuation] B --> I[Impact and conflict signals] ``` ## What You Get | Need | Rta-Smriti capability | | --- | --- | | Start a new agent task without retelling everything | Bounded context packs and structured continuation checkpoints | | Know whether recalled context deserves trust | Evidence provenance, hashes, freshness, and `pramana` labels | | Understand a repository beyond keyword search | Files, symbols, imports, calls, tests, memories, and evidence graph | | Detect drift and contradiction | Canonical project identity, bitemporal truth, conflict and decision-debt views | | Preserve long-running agent work | Incremental, redacted, resumable session capture and immutable event history | | Operate local services without guesswork | Preview-confirmed lifecycle plans, independent health axes, repair, and receipts | | Collaborate without uploading a whole project brain | Optional encrypted scopes, peer permissions, offline sync, review bundles, and audit history | | Use multiple AI coding hosts | Local stdio MCP plus recipes for Codex, Claude Code, Cursor, Zed, OpenCode, and Gemini CLI | | Keep project data private | Local SQLite, no telemetry, no cloud database, explicit capture grants | ## See The Product  <table> <tr> <td width="50%"><img src="launch-assets/screenshots/operator-graph-v1.1.0.png" alt="Rta-Smriti v1.1 repository and evidence graph" /></td> <td width="50%"><img src="launch-assets/screenshots/operator-lifecycle-v1.1.0.png" alt="Rta-Smriti v1.1 Trusted Lifecycle Supervisor controls and independent health axes" /></td> </tr> <tr> <td><strong>Inspectable project graph</strong><br />Navigate code, memories, evidence, and their relationships.</td> <td><strong>Trusted lifecycle</strong><br />Inspect, preview, apply, verify, repair, and retain a sealed receipt.</td> </tr> </table>  **Governed federation:** selected project memory is shared through explicit scopes and device permissions. The screenshot deliberately includes an offline relay and quarantined event to show recovery state rather than an all-green marketing mock. The [60-second v1 product demo](launch-assets/product-hunt/rta-smriti-v1.0.2-product-demo.mp4) was captured from `v1.0.2`. It demonstrates the Project Reality foundation; the `v1.1` Trusted Lifecycle Supervisor shown above was added later. ## How It Works ### 1. Build a local project brain Rta-Smriti indexes the canonical Git checkout into project-scoped SQLite. It records content hashes, symbols, imports, evidence, temporal state, and durable memories without uploading the project to a hosted service. ### 2. Separate evidence from memory ```mermaid flowchart TB O[Direct observation<br/>pratyaksha] --> T[Accepted project truth] S[Trusted instruction<br/>sabda] --> T I[Inference<br/>anumana] --> R[Review required] M[Prior memory<br/>smriti] --> R K[Hypothesis<br/>kalpana] --> R R -->|verified and promoted| T ``` A test result, a human instruction, an inference, and a brainstorm are not treated as the same kind of fact. Claims retain their source, verification state, and valid time. ### 3. Compile only the context the next task needs The context compiler selects direct evidence before low-trust history, obeys explicit token budgets and privacy grants, and emits a selection receipt explaining what was included and why. Search results report task-anchor relevance. If a named project or product anchor is missing, the quick context pack abstains and suppresses weakly overlapping evidence until the operator selects or registers the intended project brain. CLI and MCP context retrieval use validated query-only SQLite connections. Watcher, capture, and continuity workers enter one FIFO writer queue per brain. Queue tickets are atomically published and owner-private, stale tickets from terminated workers are removed, and the active operating-system lease is released automatically after a crash. Expired or cancelled waiters cannot enter the critical section. Repository and Codex session discovery run before their workers join the write turn; shutdown drains use a bounded final wait, while heartbeats remain responsive and oversized WALs are checkpointed in bounded turns. Optional local-model continuity compaction also runs outside the writer turn, then reacquires a bounded turn only to commit its unverified derived result. Shutdown preserves the deterministic checkpoint without waiting on optional model inference. An active SQLite transaction remains atomic and is not preempted. This keeps retrieval available during background learning without treating a read as a hidden database write. Integrations can assert that boundary explicitly with `search --json --read-only`. The JSON response includes `access.mode: read_only` and `access.writes_performed: false` as a stable machine-readable contract. When byte-identical sources exist in both current code and deployment/package mirrors, retrieval selects the current source path and collapses the mirrors. Query-only semantic retrieval loads Sentence Transformer models from the local cache only, and CLI output is normalized to UTF-8 so indexed Unicode remains printable on legacy Windows consoles. ### 4. Operate through one trustworthy boundary ```mermaid flowchart LR A[Inspect] --> B[Plan] B --> C[Review risks and backups] C --> D[Approve] D --> E[Apply] E --> F[Verify] F --> G[Immutable receipt] F -->|degraded| H[Repair or rollback] H --> F ``` Database, repository, capture, continuation, MCP, and federation health remain independent. A running process alone never proves that continuation is ready. ### 5. Share only governed project memory ```mermaid flowchart LR A[Local owner brain] -->|signed encrypted events| R[Opaque relay] R -->|authorized scopes only| B[Local peer brain] A -->|encrypted invitation| B B -->|offline changes retain provenance| R ``` Federation is off by default. Operators explicitly create scopes, verify device fingerprints, grant capabilities, preview mutations, and rotate keys after revocation. Concurrent evidence and disagreement remain visible until an explicit decision. ## Ten-Minute Start **Requirements:** Python 3.11 or newer and Git. Node.js is needed only to modify the dashboard or website source. The **v1 Project Reality CLI** is the shared entry point for onboarding, search, continuation, lifecycle operation, MCP configuration, and the local console. ### Windows PowerShell ```powershell git clone https://github.com/sulabhdubey/rta-smriti-brain.git cd .\rta-smriti-brain python -m venv .venv & .\.venv\Scripts\python.exe -m pip install . $RtaBrain = Join-Path $PWD ".venv\Scripts\rta-brain.exe" $BrainDir = "$env:USERPROFILE\Documents\Rta-Smriti\brains" & $RtaBrain start C:\path\to\my-project --project my-project --brain-dir $BrainDir --write-agents ``` <details> <summary><strong>macOS or Linux</strong></summary> ```bash git clone https://github.com/sulabhdubey/rta-smriti-brain.git cd rta-smriti-brain python3 -m venv .venv ./.venv/bin/python -m pip install . RtaBrain="$PWD/.venv/bin/rta-brain" BrainDir="$HOME/.local/share/rta-smriti/brains" "$RtaBrain" start /path/to/my-project --project my-project --brain-dir "$BrainDir" --write-agents ``` </details> `start` detects the canonical Git root, creates or migrates the brain, indexes the repository, starts managed sync, starts matching Codex continuity capture when local sessions exist, and opens an authorized local console. Use `--no-continuity` on machines without local Codex sessions. Prefer a standalone binary? Download the Windows, macOS, or Linux artifact and its SBOM from the [`v1.1.0-alpha.6` release](https://github.com/sulabhdubey/rta-smriti-brain/releases/tag/v1.1.0-alpha.6), then verify it against `SHA256SUMS.txt`. ## Private By Default ```mermaid flowchart LR subgraph Machine[Your machine] P[Project checkout] --> R[Rta-Smriti] A[Authorized agent sessions] --> R R --> S[(Local SQLite brain)] R --> C[Loopback operator console] R --> M[Local MCP server] end R -. no telemetry .-> X[No Rta-Smriti cloud] ``` - Capture is opt-in, source-scoped, bounded, and redacted before durable queuing. - Captured text remains untrusted evidence until verified or promoted. - The console binds to loopback and uses short-lived session capabilities. - Project names, paths, transcript content, databases, keys, and snapshots must not be published. Public diagnostics use bounded counts and fingerprints. - Signed or encrypted snapshots are private backup artifacts, not publication formats. Read the complete [security and privacy policy](SECURITY.md) before handling sensitive repositories. Report vulnerabilities through the private process documented there. ## Architecture At A Glance ```mermaid flowchart TB CLI[CLI] --> K[Project Cognition Kernel] MCP[stdio MCP] --> K UI[Local Operator Console] --> K K --> ID[Canonical identity] K --> EV[Append-only event journal] K --> BT[Bitemporal truth] K --> RG[Repository graph] K --> CP[Governed context compiler] K --> LS[Trusted lifecycle supervisor] K --> GF[Optional governed federation] ID --> DB[(Project-scoped SQLite)] EV --> DB BT --> DB RG --> DB ``` Rta-Smriti is deliberately an evidence and continuity layer. It does **not** execute project work, choose models, silently edit host configuration, or replace an agent harness such as RTA-Net AI. ## Release Evidence The current prerelease includes: - cross-platform CI and native binaries for Windows, macOS, and Linux; - CycloneDX SBOMs and a signed workflow provenance trail; - anonymous-download checksum verification; - installed-package, CLI, dashboard, lifecycle, MCP, privacy, and security checks; - encryption, signature, revocation, conflict, relay, recovery, and sustained-sync checks; - an explicit record of host recipes versus hosts exercised live. Green tests are not presented as proof of universal correctness. Exact scope, known limits, and remaining external-host gates are recorded in [Release Verification](docs/RELEASE_VERIFICATION.md). ## Documentation | Start here | What it covers | | --- | --- | | [Installation](docs/INSTALLATION.md) | Source install, native binaries, upgrades, troubleshooting, uninstall | | [10-minute Atlas path](docs/ATLAS_10_MINUTE_PATH.md) | Small end-to-end trial on a synthetic project | | [Usage guide](docs/USAGE_GUIDE.md) | CLI workflows, capture, checkpoints, context, snapshots, workspaces | | [Governed federation](docs/FEDERATION_GUIDE.md) | Identity, scopes, invitations, encrypted sync, review bundles, recovery | | [Architecture](docs/ARCHITECTURE.md) | Identity, event journal, truth model, graph, compiler, lifecycle | | [MCP host matrix](docs/MCP_HOST_MATRIX.md) | Host recipes, capability profiles, and live-proof status | | [Public benchmark](docs/PUBLIC_BENCHMARK.md) | Reproducible retrieval harness and honest interpretation | | [Release notes](docs/RELEASE_NOTES_v1.1.0-alpha.6.md) | Consolidated reliability and first-project pilot | | [Release verification](docs/RELEASE_VERIFICATION.md) | Tests, artifacts, checksums, security evidence, and limits | | [Contributing](CONTRIBUTING.md) | A practical first contribution path | ## Community ### Share What Happened Tried Rta-Smriti on a real project? - [Share your experience](https://github.com/sulabhdubey/rta-smriti-brain/discussions) - [Ask for installation help](https://github.com/sulabhdubey/rta-smriti-brain/discussions/categories/q-a) - [Report a bug](https://github.com/sulabhdubey/rta-smriti-brain/issues/new?template=bug_report.yml) - [Pick a first contribution](https://github.com/sulabhdubey/rta-smriti-brain/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) - [Star the repository](https://github.com/sulabhdubey/rta-smriti-brain) if it saved you from repeating project context Independent coverage: - [Harness Institute: Rta-Smriti keeps agent memory local](https://www.harness.institute/research/rta-smriti-brain-keeps-agent-memory-local) - [The Next New Thing: Rta-Smriti segment](https://www.youtube.com/watch?v=AWzzmrCPe-A&t=1350s) ## Provenance And License Conceived and researched by [Sulabh Dubey](https://github.com/sulabhdubey). Built with [OpenAI Codex](https://openai.com/codex/) as the primary design, engineering, testing, and documentation agent under Sulabh's product direction and release approval. See [Contributors](CONTRIBUTORS.md) for the full provenance statement. Released under the [MIT License](LICENSE).