Multi-agent coordination2026Active

Smriti

Version control for project reasoning state across coding agents.

Back to projectsWebsiteGitHub

At a glance

Problem

Long-running reasoning becomes brittle when usable state only exists inside transcript text.

Approach

Version, branch, compare, and restore reasoning state while agents coordinate through claims and structured tasks.

Why it matters

Agents can share current project state without treating a transcript or handoff file as the source of truth.

System sketch

01Checkpoint

Record decisions, assumptions, tasks, open questions, and repository state as one structured snapshot.

02Branch

Fork from a clean checkpoint to compare alternatives without contaminating the base line of thought.

03Coordinate

Use time-bounded claims, stable task IDs, and freshness checks so agents can see overlapping work before they collide.

04Verify

Compare recorded reasoning with the live Git repository and surface branch, commit, worktree, and dirty-state drift.

Design notes

The handoff is structured state

A current project state with stable objects carries more usable context than another handoff document.

Coordination should remain advisory

Claims make intent visible without turning Smriti into a scheduler or lock manager.

Recorded state must admit drift

A checkpoint is useful only if the system can show when the codebase has moved past it.

Question

I built Smriti after running Claude Code and Codex on the same projects. Each agent could write code, but neither could see the other agent's decisions, assumptions, or active work.

Handoff files helped with simple continuation. They broke down when work branched, two agents started near-simultaneously, or the repository moved past the recorded state.

Approach

  • Store decisions, assumptions, tasks, open questions, and artifacts as versioned checkpoints.
  • Fork, compare, and restore reasoning branches without rewriting the base state.
  • Expose active claims, stable task IDs, freshness checks, and repository drift through the CLI and MCP.
  • Keep core coordination local and deterministic; use an LLM only for optional extraction, review, and chat features.

Open questions

  • What checkpoint granularity preserves enough context without encouraging constant bookkeeping?
  • Which coordination signals should remain advisory, and which need stronger enforcement?
  • How should reasoning state evolve when agents work across repositories and longer time horizons?

Project spec

RoleConcept, architecture, and implementation.
StatusActive
TypeMulti-agent coordination
UpdatedMay 2026
Primary objectStructured checkpoints for decisions, assumptions, tasks, and open questions.
System boundaryCLI, 21 MCP tools, agent skill packs, and a human-facing UI.
Current artifactLocal-first SQLite workflow, with Postgres for shared team use.
Pressure pointKeeping shared state fresh without turning coordination into overhead.

Topics

multi-agent systemsreasoning statecoordinationMCP

Where it is now

  • The public site and repository include a guided quickstart that runs without API keys.
  • Smriti is dogfooded across active Claude Code and Codex projects.
  • The current system spans local and shared backends, 21 MCP tools, agent skill packs, repository drift checks, claims, worktrees, and a human-facing UI.