Local agent memory

Link gives every agent the same memory.

Source-backed Markdown memory for Codex, Claude, Cursor, Kiro, VS Code, Copilot, and any MCP client. Local files. Inspectable sources. Budgeted context.

Try Link View on GitHub MCP Registry

The product

Not another notes app. A local memory layer agents can actually use.

Link turns raw notes, transcripts, project context, and explicit memories into a source-backed wiki. Agents query a compact packet instead of reading your entire folder.

Personal memory

Preferences, decisions, facts, and project context stay durable across agent sessions.

Source-backed wiki

Raw sources compile into Markdown pages with citations, backlinks, and reviewable provenance.

MCP-native recall

One local memory works across MCP-capable tools through the same query, brief, graph, and memory lifecycle tools.

Budgeted context

Smart query packets return the right memory, pages, graph neighborhood, and follow-up actions without flooding tokens.

Private by default

No hosted backend, no telemetry, no cloud lock-in. Your memory stays on disk as plain Markdown.

Auditable lifecycle

Capture, propose, approve, review, archive, restore, forget, and explain what Link remembers.

First 2 minutes

Run a finished memory wiki locally.

The demo includes raw sources, wiki pages, one starter memory, backlinks, graph context, and a compact query packet.

First 10 minutes Troubleshooting 
git clone https://github.com/gowtham0992/link.git
cd link
python3 link.py demo
python3 link.py serve link-demo

# then try
python3 link.py query "why does Link help agents?" link-demo --budget small
python3 link.py brief "working on agent memory" link-demo
python3 link.py benchmark "agent memory" link-demo

Visible trust

Memory is inspectable, reviewable, and explainable.

A memory is not a hidden vector. It is a Markdown page with status, scope, source, review state, graph links, and an audit trail.

Link memory dashboard
Memory dashboard: profile, inbox, captures, and review states.
Link explain memory view
Explain memory: why Link knows something and whether it is ready to recall.
Link graph view
Graph view: bounded by default, expandable when you need the whole neighborhood.

Agent contract

Agents get a small set of reliable moves.

Check readiness, brief before work, ingest raw files, remember explicit facts, query smart context, validate after writes, and explain why a memory exists.

BeforeRepeated context in every chat.
AfterShared local memory across agents.
StoragePlain Markdown and JSON indexes.
ScaleSQLite FTS, bounded graph payloads, local cache.

Read next

Start small, then make it your agent memory.

The docs are arranged by user path: try the demo, understand the model, wire MCP, then use the CLI and maintenance tools when you need them.

First 10 minutes

Run the demo, add one source, save one direct memory, and verify the loop.

Concepts

Understand raw sources, wiki pages, memories, graph indexes, and budgeted query packets.

MCP setup

Install the MCP server and teach local agents how to use Link reliably.

CLI reference

Every local command, grouped by daily workflow and maintenance jobs.

HTTP API

Local endpoints for status, query, memory, graph, validation, and web UI actions.

Security model

Local-first constraints, secret scanning, backup behavior, and HTTP safety boundaries.

Contributing

PR expectations, test gates, branch policy, and what not to include in public changes.

Troubleshooting

Fix MCP setup, blocked ingest, stale graph indexes, slow wikis, and Python packaging issues.