AgentRoots

Different agents. Same roots.

Evidence-governed continuity for agents that explore, build, and research together.

Built by Shanmukha Vellamcheti and OpenAI Codex.

Agents make exploration dramatically faster, but temporary contexts make useful work disposable. Files get reread, failed paths get repeated, facts blur, and every fresh agent must reconstruct the project's state. AgentRoots gives agents one durable, reviewed frontier instead.

Memory preserves experience. AgentRoots governs the frontier.

AgentRoots is an open-source Agent Continuity MCP for Codex, Claude, DeepSeek, and generic MCP clients. It preserves where a project came from, what is currently supported, and what should happen next, without replaying transcripts or duplicating large artifacts.

One agent can preserve useful state across sessions. Multiple agents can propose, review, and reuse the same findings. AgentRoots does not spawn, schedule, route, or execute agents.

See project state at a glance

This synthetic overview exercises all 14 record types and all 12 relationship types in the current contract. The human-readable graph is generated from the same versioned event ledger that agents query, making the complete project model inspectable without creating a second source of truth.

Related MCP server: trw-mcp

Why I built AgentRoots

My research requires exploring many hypotheses and experimental paths. Before coding agents, the number of experiments I could run manually was naturally limited. Agents changed that. They made hypothesis exploration and project-state growth dramatically faster, but they also created a new memory-management problem.

Modern agentic work increasingly depends on orchestrators and subagents for speed and cost efficiency. That can multiply duplicated work. If an orchestrator assigns two independent tasks in the same codebase, both subagents may reread the same files to understand the project, after the orchestrator already read them to make the plan. The same knowledge may be reconstructed three times. Across longer projects, context gets mixed, facts blur, failed paths are repeated, and the latest working frontier becomes difficult to recover.

Memory tools preserve experience. Planning tools preserve intent. Provenance tools preserve what ran and changed. AgentRoots connects those concerns as evidence-governed project state: what is the project's origin, what is currently accepted, what evidence supports it, what is stale or disputed, which goals remain active, and what questions or experiments should happen next. This lets agents across models and harnesses inherit a compact, grounded frontier, then verify only what their task requires instead of rebuilding the entire context from scratch.

I built AgentRoots because I needed agents to share more than memories. I needed them to inhabit the same evolving state, avoid duplicated exploration, and continue from the real frontier.

Why AgentRoots

AI agents are temporary. Their work should not be. AgentRoots preserves goals, questions, hypotheses, experiments, observations, findings, decisions, failures, and evidence across sessions, models, and harnesses.

Branches explore. Roots remember.

Quick start

python -m pip install "agentroots @ git+https://github.com/shanmukha-here/agentroots.git"
agentroots propose demo hypothesis "Caching helps" "Latency should fall." --actor codex
agentroots-mcp

AgentRoots is not published to PyPI yet. Contributors cloning the repository can instead use python -m pip install -e .. Python 3.11 or newer is required.

State defaults to the OS or XDG user data directory. Override it with AGENTROOTS_DB or --db. The legacy RESEARCH_STATE_DB variable remains accepted for local migration. SQLite runs in WAL mode. Generated state stays outside the repository.

MCP surface

Tools: research_get_context, research_get_frontier, research_query, research_get_record, research_get_graph, research_propose, research_revise, research_review, research_link_evidence, research_mlflow, research_sync, and research_validate.

Resources: project brief, project frontier, record, and context packet under the research:// URI scheme. Protocol names remain research-specific because the initial ontology models evidence-backed investigative work. AgentRoots branding covers its broader engineering, research, and long-running agent uses.

Exact resource templates:

• research://project/{project}/brief

• research://project/{project}/frontier

• research://record/{record_id}

• research://packet/{packet_id}

Example MCP argument shapes:

{"tool":"research_propose","arguments":{"project":"demo","record_type":"finding","title":"Cache result","body":"The cache reduced repeated reads by 12 percent in the measured workflow. The comparison used the same task fixture and code revision. This supports retaining the cache for subsequent trials. A replication should confirm the result on a larger repository.","creator":"codex"}}
{"tool":"research_review","arguments":{"record_id":"UUID","actor":"reviewer","verdict":"accepted","resolves_record_ids":["GOAL_UUID"]}}
{"tool":"research_link_evidence","arguments":{"record_id":"UUID","uri":"mlflow://runs/123","kind":"mlflow-run","actor":"reviewer","content_hash":"sha256-if-known"}}
{"tool":"research_get_context","arguments":{"project":"demo","query":"cache","token_budget":1500}}
{"tool":"research_mlflow","arguments":{"operation":"link","record_id":"UUID","run_id":"RUN_ID","actor":"reviewer","include_artifacts":true}}

research_sync imports supplied events, exports current project events, and can mark packet record IDs as used. CLI export and import provide file-based JSONL transfer.

CLI query text is positional. Run agentroots <command> --help for command-specific arguments:

agentroots context demo "cache latency" --tokens 1500
agentroots validate demo
agentroots export demo events.jsonl
agentroots graph demo project-map.html

The graph command creates a self-contained, read-only React Flow knowledge map. It works offline and supports automatic layouts, searching, lifecycle and type filters, pan and zoom, evidence inspection, version metadata, relationship tracing, and copying record IDs for review or correction. See graph viewer architecture for customization and the governed editing roadmap.

On Windows, prefer these positional CLI commands or MCP tool calls over hand-escaped JSON in PowerShell. For contributor tests in a clean checkout:

python -m pip install -e ".[dev]"
python -m pytest

Lifecycle: candidate to provisional to accepted, plus disputed, rejected, superseded, and stale. Creators cannot accept their own proposals by default. Acceptance requires resolvable evidence. Mutations emit append-only events. Stored text is always treated as untrusted data. Substantive records should normally explain context, evidence, implications, and next steps in three to five sentences. research_validate warns about thin provisional or accepted records; set metadata.concise_fact=true only when a shorter statement is genuinely complete. An accepted finding can explicitly resolve one or more goals. A resolves link removes those goals from the active frontier while preserving their full history. supports does not close a goal.

Implemented today:

• SQLite event ledger, revisions, projections, FTS5, and fuzzy lookup

• sectioned, token-budgeted, audited context packets

• review governance, contradictions, failed-attempt recall, and Git staleness

• exact JSONL event sync plus backup and restore

• read-only MLflow evidence integration and Trackio adapter

• H-E-F and signac importers

• stdio MCP server, CLI, schemas, tests, fixtures, and three-agent demo

Flowcept, AiiDA, PostgreSQL, remote HTTP, ACLs, and UI remain roadmap work. See the specification, architecture, integrations, roadmap, evaluation, and demo.

Contributors

See CONTRIBUTORS.md. Contributions are welcome under Apache-2.0.