sessionweave 0.1.0

Local-first context memory engine for AI-agent workflows.

English

A local memory layer for AI coding agents.

SessionWeave gives Codex, Claude Code, Cursor, Gemini, GitHub Copilot, and ChatGPT persistent project memory that survives across sessions.

Instead of repeating the same decisions, constraints, rejected ideas, and implementation context every time, you record durable memory once. SessionWeave stores it locally in .sessionweave/, builds a lightweight memory graph, and routes only the relevant context back to your coding agent.

AI coding agents forget. SessionWeave gives them repo-local memory.

sessionweave install --platform codex
sessionweave start-session --id demo
sessionweave route --task "continue the auth refactor"

.sessionweave/
├── graph.db                 local memory graph
├── journal/                 append-only session journals
├── pages/context-packet.md  compact routed context for the current task
├── commits/                 Git-like memory commits
└── instructions/            adapter instructions for coding agents

Print a copy-pasteable first-run guide:

sessionweave quickstart
sessionweave quickstart --output SESSIONWEAVE_QUICKSTART.md

What SessionWeave Solves

AI coding agents are powerful, but they lose project continuity.

They forget:

• why a decision was made
• what ideas were rejected
• which files matter for a task
• what happened in the previous session
• what context should not be revived

SessionWeave fixes that with explicit, local, task-routed memory.

How It Works

SessionWeave runs as a local CLI in your repo.

1. Start a session.
2. Record important decisions and constraints.
3. Route compact context for the current task.
4. Close the session to create a memory commit.
5. Your next agent session starts from durable project memory, not raw chat history.

sessionweave start-session --id feature-x
sessionweave record --id feature-x --role user --text "<sessionweave-memory block>"
sessionweave route --task "continue feature-x implementation"
sessionweave close-session --id feature-x --summary "Finished routing layer"

Strict Memory, Not Guesswork

SessionWeave does not infer durable memory from random prose.

Durable memory is explicit:

{
  "nodes": [
    {
      "type": "decision",
      "title": "Use local SQLite storage",
      "summary": "Keep project memory local-first and dependency-light.",
      "temperature": "hot",
      "importance": 5
    },
    {
      "type": "anti_memory",
      "title": "Do not dump full chat history",
      "summary": "Route compact task context instead of loading raw sessions.",
      "temperature": "pinned",
      "importance": 5
    }
  ],
  "edges": []
}

This keeps memory auditable and prevents the agent from inventing project history.

Platform Support

Platform	Generated files
Codex	AGENTS.md, .sessionweave/instructions/codex.md
Claude Code	CLAUDE.md, .sessionweave/instructions/claude-skill/SKILL.md
Cursor	.cursor/rules/sessionweave.mdc
Gemini CLI	GEMINI.md, .sessionweave/instructions/gemini.md
GitHub Copilot Chat	.github/copilot-instructions.md
ChatGPT	.sessionweave/instructions/chatgpt.md

Install all adapters:

sessionweave install

Install only Codex:

sessionweave install --platform codex

Real-World Smoke Test

SessionWeave was tested on a real django/django checkout.

Initial after install:
nodes=0, edges=0, sessions=0, commits=0

After recording Django URL resolver memory:
nodes=4, edges=7, session=open

After linking real code references:
nodes=9, edges=11, session=open

After route + close-session:
nodes=9, edges=11, sessions=closed=1, commits=1, uncommitted=no

The routed context selected relevant Django URL resolver memory and kept anti-memory separate instead of loading the whole repository.

graph TD
  S["Session: django_codex_real"]
  P["Django real-world codebase"]
  D["Use Codex adapter in Django checkout"]
  T["Investigate Django URL resolver workflow"]
  A["Do not dump the whole Django repository"]
  F1["file: django/urls/resolvers.py"]
  F2["file: tests/urlpatterns_reverse/tests.py"]
  F3["file: docs/topics/http/urls.txt"]
  C["commit: 017d7f6"]

  S --> P
  S --> D
  S --> T
  S --> A
  P --> D
  P --> T
  T --> A
  T --> F1
  T --> F2
  T --> F3
  T --> C

Diagnostics

SessionWeave ships with release and local health checks:

sessionweave doctor
sessionweave lint
sessionweave dogfood
sessionweave release-check

Install

After public release:

uv tool install sessionweave
sessionweave install

or:

pipx install sessionweave
sessionweave install

or:

npm install -g sessionweave
sessionweave install

From source:

git clone https://github.com/adibisoyi/SessionWeave
cd SessionWeave
python -m pip install -e .
sessionweave install

Privacy

SessionWeave is local-first.

• Memory is stored in .sessionweave/.
• No hosted service is required.
• No telemetry.
• No vector database required.
• No model account modification.
• Adapter installers write local instruction files only.

Current Limits

• SessionWeave is local-first alpha software.
• It does not run an always-on daemon.
• It does not provide hosted sync, team permissions, or enterprise governance.
• Adapter installers are local instruction generators, not official marketplace integrations.
• Durable memory must be explicit structured memory.

Current Status

SessionWeave is public-alpha software.