MathResearch Ledger

Durable structured memory and verification infrastructure for long-horizon mathematical research across external AI coding and research harnesses.

MathResearch Ledger is not another research-agent runtime. External agents do the research; Ledger preserves durable research state. Use Codex or another harness for reasoning, reading and experiments, then carry the same research workspace into the next session without reconstructing it from chat.

中文说明 · Changelog

1.2.2: eleven tools, including Problem lifecycle, Codex session-workspace bundled MCP and DSH per-Agent workspace binding. See the release notes, adapter contract and measured capability matrix.

Why keep a research ledger?

A long proof may span many sessions. Chat summaries can lose assumptions, repeat failed routes or cite a claim whose dependencies have changed. Ledger keeps Problems, Claims, Gaps, proofs, counterexamples and failed attempts as structured records, along with their sources, versions and checkpoint history.

• Resume deliberately: restore an explicit Problem, its primary target, open gaps and latest meaningful checkpoint.
• Retrieve selectively: use exact identifiers, bounded context and paginated originals; preserve full research history without loading it all into every conversation.
• Track what a result relies on: qualified identities and pinned dependency versions distinguish current, stale and unknown premises.
• Checkpoint meaningful progress: preview a Research Episode, retain its exact source and publish provisional records while keeping next directions advisory.
• Keep evidence separate from authority: verification, review, canonicalization and human acceptance have distinct roles.

External research agent (Codex / other harness)
                       │ eleven MCP tools
               MathResearch Ledger
             state · context · evidence
                       │
             .math-research-ledger/

Five-minute quickstart

Requires Python 3.11+. From a cloned repository:

git clone https://github.com/Caie777/math-research-ledger.git
cd math-research-ledger
python -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install -e '.[mcp]'
math-research-ledger --workspace 'C:\Research\MyProject' init
math-research-ledger --workspace 'C:\Research\MyProject' ledger create-problem 'My question' 'State the question and assumptions here.'
math-research-ledger --workspace 'C:\Research\MyProject' status

On macOS/Linux, activate with source .venv/bin/activate and use your own workspace path. Initialization is explicit: choose a research directory outside the tool checkout. It creates empty state, workspace.yaml and research AGENTS.md, preserves existing instructions and ordinary files, and creates no demonstration research objects. The create-problem command above creates the Problem you explicitly request.

For a downloaded wheel, use python -m pip install './math_research_ledger-1.2.2-py3-none-any.whl[mcp]' instead of the editable install. See GitHub Releases. Then follow Codex plugin setup, enable the plugin for your research project and start a fresh session. Installation and enablement are separate; ordinary projects can leave it disabled. A standalone MCP client can launch math-research-ledger-mcp --workspace '<research-workspace>' over stdio. MCP never initializes missing state. CLI aliases smr, math-research and smr-mcp remain available; the Python import namespace is math_research.

A minimal continuation workflow

Free research by default; memory intervention by exception.

1. Select the existing Problem explicitly with research_status(problem_id="PROBLEM-0001"); the MCP server must already be pinned at startup. Inspect the primary target, gaps, risks, latest checkpoint and advisory handoff.
2. Read the exact target with research_get_object; restore context once with research_context_packet(ref="<qualified target>", purpose="CONTINUE_RESEARCH", token_budget=8000). Check omissions, truncation, freshness and premise eligibility.
3. Research freely. Retrieve again when switching a major route, reusing a key premise, suspecting an old failed attempt or resolving a conflict with history.
4. At meaningful progress, prepare a Research Episode. Separate Confirmed Research Increment from advisory Next Research Directions. Preview with research_ingest_episode(dry_run=true), resolve ERRORs, understand WARNINGs, then ingest once when authorized. The next session restores from the Ledger.

These IDs illustrate the call pattern: use your actual Problem and qualified refs. Search hits are candidates, not trusted premises. Inspect exact objects, input versions, provenance and evidence before relying on them. Preserve failed routes and full proofs; do not persist every chat turn or rebuild context after every local deduction.

Eleven MCP tools

Tool	Use
research_create_problem	Preview/register a new active Problem
research_update_problem	Preview metadata, lifecycle or explicit semantic revision
research_status	Restore Problem state, continuation navigation and counts
research_search	Find bounded relevance candidates or an exact identity
research_get_object	Read exact objects, provenance and paginated originals
research_context_packet	Restore budgeted context for a target
research_review_packet	Prepare version-pinned material for advisory review
research_stale_report	Inspect dependency/version drift
research_ingest_episode	Preview or ingest a provisional research checkpoint
research_save_note	Capture an important isolated provisional increment
research_request_acceptance	Request human acceptance; never perform it

Episode Markdown is limited to 128,000 characters per MCP call. The lower-level file ingestion ceiling and bounded responses are described in the Episode contract. No shell, arbitrary file-write, compiler execution or human-accept operation is exposed by these tools.

Trust and workspace ownership

checked != accepted; canonical != verified. Soft review is advisory. Freshness is derived current state; historical verification and acceptance receipts remain historical. Ingested output stays provisional, and source-reported revisions do not automatically revise, refute or invalidate canonical objects.

Human acceptance is not exposed through the model-callable MCP surface and is enforced as a workflow boundary. It is not an authentication or sandbox boundary against arbitrary local code or filesystem access. The human CLI is a trusted channel; see architecture and authority and verification.

The harness owns the workspace; the user owns ordinary files. Ledger owns only .math-research-ledger/. Keep that directory, workspace.yaml and research AGENTS.md local-only. Ledger never stages, commits, pushes or chooses research for publication. An optional notes repository contains only work the user chooses to publish. Workspace format and migration explains ownership and Git-local exclusions. Historical provenance is not rewritten by tool upgrades.

Integrations and limitations

• Codex: prepare the Codex-native plugin with a persistent runtime; one plugin ON/OFF controls its shared skill and bundled MCP. The server binds exactly to the Host's startup session cwd using --workspace ., without parent discovery or auto-init. Installation and configuration. This is tested Codex 0.160.0 behavior, not a guaranteed project-root API. See the acceptance matrix.
• DeepSeek Harness: the Agent Host plugin binds each new Agent to its immutable session root through the official MCP client. Windows DSH 0.2.1-rc.3 isolated Host and real Profile read-only cutover passed, including eleven tools, workspace isolation, resume and OFF/ON. The configuration-only Bundle remains a manual / single-workspace fallback. Do not enable both at once. rc.2 is unverified for this adapter; GUI conversational acceptance is not claimed.

Known limitations: DSH macOS/Linux and formal writes to a real research corpus have not been separately validated. Each MCP session has one pinned workspace. The plugin does not change permissions or Full Access. GitHub publication does not imply a public Plugin Directory listing.

Development

Use separate synthetic temporary workspaces; do not initialize research state here.

python -m pip install -e '.[dev,mcp]'
python -m pytest
ruff check src tests scripts benchmarks
ruff format --check src tests scripts benchmarks
python -m mypy
python -m build
git diff --check

Architecture · Research assets · Historical engineering reports