Skip to main content
Glama
Sign In
deenesjakoruzh
Perenna

Perenna

by scarletkc
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Hybrid

Perenna

Python PyPI CI Codecov CodeRabbit Pull Request Reviews License Ask DeepWiki perenna MCP server

A lightweight, Git-backed permanent memory for AI agents. Claude Code, Codex, ChatGPT, Cursor, and other MCP clients can share durable memories without sharing a vendor account or conversation history.

  • Separate MCP tools for reading, writing, and deleting memories

  • Local stdio, loopback-only Streamable HTTP, and OAuth-protected remote HTTP

  • Human-readable Markdown stored in an independent Git repository

  • Local Vexor retrieval index that can always be rebuilt from Git

  • Cross-process locking for multiple local agent processes

Why Perenna?

Your memory should follow you, not the agent you happen to be using.

Claude Code, Codex, Cursor, and ChatGPT keep memory in separate silos. Switch agents and your memory disappears. Switch machines and local memory stays behind.

Perenna gives them one shared, Git-backed memory. Local agents and ChatGPT can connect to the same self-hosted Perenna service, while every durable memory stays ordinary Markdown you can inspect, edit, version, and back up yourself.

Perenna stays focused on permanent memory: no required account, proprietary memory cloud, or automatic conversation extraction.

Related MCP server: Hypermnesic

Quickstart

Install with your AI agent

Paste this into Claude Code, Codex, ChatGPT Desktop, Cursor, or another coding agent with terminal and local MCP configuration access:

Open the following URL, read the complete instructions, and follow them to
install and connect Perenna:
https://raw.githubusercontent.com/scarletkc/Perenna/main/docs/guides/agent-installation.md

Install a published release

Perenna requires Python 3.12+, Git, and uv.

Install Perenna

uv tool install perenna

Configure retrieval

Perenna needs a working Vexor embedding provider. For interactive provider selection and configuration, run:

uvx vexor init

Perenna automatically reuses ~/.vexor/config.json. When using process-level configuration, make sure the MCP server receives VEXOR_CONFIG_JSON plus VEXOR_API_KEY or the selected provider's key from its host environment. Remote providers receive memory text and search queries.

If you choose local embeddings, also install Perenna's local extra:

uv tool install "perenna[local]"

Vexor provider configuration covers remote and local setup. From the environment that starts the MCP client, verify the selected provider with:

uvx vexor doctor

Connect a client

For the standalone setup, configure the MCP client to start:

perenna mcp

Follow the Client setup guide for the exact client command or configuration file.

Codex and Claude Code can also install the optional memory behavior Skill:

perenna skill install --agent codex
# or
perenna skill install --agent claude-code

Repeat --agent in one command when both clients should receive the skill. The configuration reference documents user and project scope, destinations, and replacement safeguards.

Codex and Claude Code can instead install the combined Skill and MCP connection from Perenna's repository Marketplace. Follow the Plugin setup guide and choose one setup path per client.

Perenna creates its local data under ~/.perenna/ unless another home is configured.

Add optional Git synchronization

To import, publish, or fast-forward compatible history through a private Git repository, run:

perenna sync setup <repository-url>

Successful setup saves the selected remote in the Perenna home. Use perenna sync disable to return to saved local-only mode without removing the Git remote.

The configuration reference owns remote selection, credentials, conflict handling, and recovery details.

Install from source

git clone https://github.com/scarletkc/Perenna.git
cd Perenna
uv tool install .

Source contributors should use the locked environment in the development guide.

Documentation

Start with the documentation index, then follow the path for your task:

License

MIT

Available Tools

3 tools
memory_deleteA
Destructive

Delete exactly one committed memory by ID, expected title, and revision. The deletion removes it from current recall but remains recoverable from Git history.

Inputs: memory_id (required), base_revision (required), expected_title (required). Returns: action, commit, memory, changed, +3 more

memory_readA
Read-onlyIdempotent

Read permanent memory. List returns stable memory IDs and titles, search returns bounded ranked candidate passages, and get returns one complete committed memory with its revision.

Returns: object

memory_writeA
Destructive

Create or modify permanent memory. Summary is a stable one-line description of what a memory covers. Patch applies exact all-or-nothing edits; replace overwrites the complete summary and body. Existing memories require a current base revision.

Returns: action, commit, memory, changed, +2 more

Tool Schema Changelog

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

Loading tool schema changelog…

    Deploy Server
    One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

    TDQS

    A4.2/5.0

    Scored across 3 tools

    Disambiguation5/5

    Each tool handles a distinct lifecycle operation: reading/querying, writing/updating, and deleting memories. There is no overlap between the primary actions, and the sub-modes within memory_read are explicitly separated (list, search, get).

    Naming Consistency5/5

    All tool names follow the same memory_verb pattern using clear, lowercase snake_case verbs. The naming makes the operation type immediately obvious and perfectly consistent across the set.

    Tool Count5/5

    Three tools is a compact but complete surface for a permanent memory store. Each tool earns its place, and no redundant or extraneous operations exist.

    Completeness5/5

    The set covers the full lifecycle: create/modify via memory_write, read via memory_read (including list/search/get), and delete via memory_delete. Revision-handling and Git-history recovery details further round out the functionality, leaving no critical gaps.

    Maintenance

    ActivityStale
    ResponsivenessNo recent community issues

    Resources

    Related MCP Connectors

    Related MCP Servers

    • Hypermnesic

      leonardsellem
      A
      license
      A
      quality
      B
      maintenance
      Git-native long-term memory for AI agents: your markdown files are the source of truth, the search index is a disposable projection rebuilt from git, and every memory the agent writes is a reviewable git commit. Served over one OAuth-secured MCP endpoint with hybrid lexical+semantic recall and a gated, git-first commit_note write tool.
      7
      8
      AGPL 3.0

    • gcc-mcp

      yanhao13
      A
      license
      Not graded
      quality
      D
      maintenance
      Persistent, structured memory for AI coding agents via a git-like filesystem with branch/commit/merge tools, enabling agents to maintain context across sessions.
      MIT