Skip to main content
Glama
NeuroWeave Timeline

NeuroWeave Timeline

by Thatgfsj

NeuroWeave Timeline (NWT)

Process memory for AI agents and humans. NWT remembers how a project became what it is — not just what it is now.

CI License: MIT Python 3.10+ PyPI: NWtimeline MCP Release v0.2.0 GitHub stars

Most tools remember results. NWT remembers evolution.

Traditional memory:  User → Context → Summary → Memory
Timeline memory:     User → Action  → Timeline Event → Evolution Graph

Every meaningful action in your project — a decision, a refactor, a file creation, a bug fix — becomes a node in a durable timeline. The links between nodes form an Evolution Graph that explains why the project looks the way it does today.


What you can ask

Question

One-liner

Why does this file exist?

nwt explain activation.py

Why was this architecture chosen?

nwt search "architecture decision"

What happened three months ago?

nwt history

What decisions led to the current design?

nwt story

Show me the evolution graph

nwt graph

What did that commit actually do?

nwt git-hook-status

What changed between two events?

nwt diff 1 5

Merge small consecutive events

nwt compact

AI agents reach the same answers over MCP — see MCP integration.

Related MCP server: ContextWeave

Auto-grow: timeline from git commit

The single biggest reason NWT timelines are empty is friction. The v0.2 fix is a post-commit hook that logs an event on every commit — no human or agent typing required.

nwt init
nwt install-git-hook
git commit -m "Refactor retrieval layer" -m "Reason: lookup latency was high"
nwt: logged [12] Refactor retrieval layer

Add --strict to refuse events without a Reason: line, and --ai-command "claude -p" to let an external model fill the reason when the human didn't. See docs/git-hook.md.


30-second quick start

pip install -e .
cd your-project
nwt init
nwt log "Add activation engine" \
      --files activation.py \
      --reason "retrieval was slow"
nwt history
nwt graph

That's it. Storage is plain JSON under .nwt/. No database, no embeddings, no vendor lock-in, no daemon.


A tour of the output

nwt history — what happened, in order

  [1] 2026-06-15  Project scaffolded  [setup, milestone]
      reason: Kickoff the MVP
      files:  pyproject.toml, README.md
  [2] 2026-06-15  Add memory engine  [core, milestone]
      reason: Need a place to put things
      files:  memory.py
  [3] 2026-06-15  Add activation spreading  [memory, optimization]
      reason: Retrieval was sequential and slow
      files:  activation.py, retriever.py
  [4] 2026-06-15  Add decay mechanism  [memory]
      reason: Stale nodes should fade
      files:  activation.py
  [5] 2026-06-15  Vectorize activation  [refactor, performance]
      reason: Loop was the hot path in profiling
      files:  activation.py

nwt graph — the evolution as a tree

○    1  Project scaffolded
│
│    2  Add memory engine
      ├─ sibling  → [   4] Add decay mechanism
      └─ extends  → [   3] Add activation spreading
│
│    3  Add activation spreading
│
│    5  Vectorize activation

nwt story — 100 events compressed to one page

# memory-engine-demo — evolution summary

span: 2026-06-15 → 2026-06-15  (5 events)

milestones:
  - 1  Project scaffolded          — Kickoff the MVP
  - 2  Add memory engine           — Need a place to put things
  - 3  Add activation spreading    — Retrieval was sequential and slow
  - 4  Add decay mechanism         — Stale nodes should fade
  - 5  Vectorize activation        — Loop was the hot path in profiling

spine file: activation.py

decisions (events with stated reasons):
  - [1] Project scaffolded: Kickoff the MVP
  - [2] Add memory engine: Need a place to put things
  ...

nwt explain activation.py — why a file exists

# activation.py
created in:  event 3
modified in: 4, 5

reason:
  Retrieval was sequential and slow

New in v0.2

Events chain automatically

nwt log appends after the latest event by default, so a plain sequence of logs forms one linear history. Start an explicit branch with --parent none, or continue after a specific event with --parent ID.

Event importance

Events now have an importance field: low, normal (default), high, milestone.

nwt log "Fix login bug" --summary "..." --importance high
nwt log "Project launched" --summary "..." --importance milestone

The story command reports events grouped by importance.

nwt diff — compare two events

nwt diff 1 5
# Output:
# Diff: [1] → [5]
# Events: 5 between these points
# Added: new_feature.py
# Removed: old_feature.py
# In both (likely modified): main.py

nwt compact — merge small events

When you have many small consecutive events with the same tags, compact them:

nwt compact
# Compacted: 50 → 35 events (merged 15)
# backup: .nwt/snapshots/pre-compact-20260920T120000

Options:

  • --time-window 3600 — group events within 1 hour (default)

  • --min-group 3 — minimum group size to merge (default)

  • --dry-run — preview the merge without touching disk

Compact backs up timeline/, relations/ and the counter to .nwt/snapshots/ before rewriting, and remaps parents, typed relations, indices, and the id counter onto the renumbered events.


How it works

NWT lives in your project as a single .nwt/ directory:

your-project/
└── .nwt/
    ├── metadata.json       # project name, schema version
    ├── .counter.json       # next event id
    ├── timeline/           # one JSON file per event
    │   ├── 000001.json
    │   ├── 000002.json
    │   └── ...
    ├── relations/          # typed edges out of each source event
    ├── snapshots/          # pre-compact backups land here
    └── indices/            # derived, rebuildable
        ├── files.json
        └── tags.json

Everything is JSON, atomically written. The whole workspace is grep-friendly and git diff-friendly. See docs/architecture.md for the rationale.


MCP integration

For agent developers — NWT ships an MCP server exposing the same answers as tools:

Tool

Returns

create_event

A persisted event with id and timestamp

search_history

Matching events across task/summary/reason/files/tags

get_project_story

Compressed project story (milestones, decisions, spine file)

explain_file

Created/modified-in + earliest reason for a file

Wire it up in your MCP client:

{
  "mcpServers": {
    "nwt": {
      "command": "nwt-mcp",
      "env": { "NWT_ROOT": "/absolute/path/to/your/project" }
    }
  }
}

The server picks the workspace from $NWT_ROOT if set, else its own cwd. See docs/mcp.md for the recommended agent loop:

  1. Session start: call get_project_story to load context.

  2. For unfamiliar files: call explain_file rather than reading cold.

  3. As work is done: call create_event with a reason explaining why.

  4. When uncertain: call search_history with a hypothesis from the current code.


安装

# from PyPI
pip install NWtimeline

# from a clone (editable)
git clone https://github.com/Thatgfsj/neuroweave-timeline
cd neuroweave-timeline
pip install -e .

需要 Python 3.10+。CLI 依赖于 click;MCP 服务器依赖于 mcp。两者都是自动安装的。

安装开发依赖(pytest)并运行测试套件:

pip install -e ".[dev]"
pytest -q

Roadmap

v0.2 (this release) is the "auto-grow" cut: git hook integration with strict mode and AI-filled reasons, automatic event chaining, the importance field, diff, and compact. The timeline fills itself now.

  • v0.3 — snapshots (nwt snapshot / nwt restore) and deeper git linkage (commit SHAs on every event, nwt rewind).

  • v0.4 — multi-agent collaboration history (concurrency-safe writes, per-author identity).

  • v0.5 — NWC integration, only if NWT earns it on its own.

See docs/roadmap.md and docs/standalone.md for the full story.


Contributing

Issues and PRs are welcome. The whole project is ~1,500 lines of Python plus docs — easy to read end-to-end. Start with docs/architecture.md for the layout and CONTRIBUTING.md for the workflow.


Security

NWT stores only what you give it, on disk, in your project's .nwt/. It does not phone home, does not read environment variables other than NWT_ROOT, and writes nowhere else. The .gitignore refuses to track tokens, keys, or .env files. See SECURITY.md for the full policy.


License

MIT — see LICENSE.

Available Tools

4 tools
create_eventA

Append a new event to the project's timeline.

Args: task: Short imperative title (e.g. "Add activation engine"). summary: One or two sentences describing what was done. reason: Why this change was made. Strongly encouraged — this is what turns NWT from a log into a history. files: Project-relative file paths this event touched. tags: Free-form labels (e.g. ["memory", "optimization"]). parent: Id of the preceding event in the linear chain, or null to start a new branch.

Returns: The persisted event as a dict (including its allocated id and timestamp).

Inputs: task (required), summary (required), tags, files, +2 more. Returns: object

explain_fileA

Explain why a file exists.

Args: file_path: Project-relative path to the file (e.g. activation.py).

Returns: A dict with keys file, created_in, modified_in, events, reason and a human-readable text rendering.

Example: >>> explain_file("activation.py") { "file": "activation.py", "created_in": 23, "modified_in": [45, 67], "reason": "Improve graph retrieval performance.", "text": "activation.py\n created in event 23\n ..." }

Inputs: file_path (required). Returns: object

get_project_storyA

Return a compressed project story.

The story contains: * project name and first/last event timestamps * up to max_milestones milestone events * the file that changed most often (the "spine file") * all decision events (those with a non-empty reason)

Returns: A dict mirroring the structure of :class:ProjectStory, plus a human-readable text rendering for direct LLM consumption.

Inputs: max_milestones. Returns: object

search_historyA

Search the timeline.

Args: query: Substring to search for (case-insensitive). Matched against task, summary, reason, file paths, and tags. limit: Cap on the number of results. search_files: Include file paths in the search. search_tags: Include tags in the search.

Returns: A list of matching events, ordered by id ascending. Each event is a dict; see create_event for the field set.

Inputs: query (required), limit, search_tags, search_files. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clear, distinct purpose: creating events, explaining a file, retrieving a project story, and searching history. There is no overlap or ambiguity.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern in snake_case (create_event, explain_file, get_project_story, search_history), making it easy to predict function.

    Tool Count5/5

    Four tools cover the core functionalities of a timeline system (creation, explanation, overview, search) without excess or deficiency.

    Completeness4/5

    The set covers creation and querying well, but lacks direct event retrieval by ID and update/delete operations. However, these may be intentionally omitted for an immutable timeline.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      C
      maintenance
      Enables AI coding agents and IDEs to capture conversations, extract structured knowledge into a queryable graph database, and recall past decisions, patterns, files, and errors across all projects and workspaces. Provides persistent memory tools for saving, searching, and cleaning up knowledge via MCP.
      2
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Provides persistent memory for AI coding agents, enabling them to save, search (hybrid vector/BM25/concept graph), list sessions, and forget memories via MCP tools.
      11 npm
      Apache 2.0