[FEATURE] Community proxy that adds persistent memory, smart compaction, cache keepalive and token optimization to Claude Code

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [ ] This is a single feature request (not multiple features)

Problem Statement

Problem Statement

Claude Code has several limitations that compound each other during
real development work:

• No memory between sessions — every time you close the terminal,

Claude forgets your stack, your decisions, your conventions.
You re-explain everything from scratch every session.

• Blind compaction — context compaction fires at a fixed percentage

and summarizes everything into a lossy blob. Schema decisions,
variable names, architectural rationale get dropped. Post-compaction
Claude contradicts its own prior choices.

• Token bloat — Claude Code sends approximately 55,000 tokens of tool

definitions on every request regardless of what the turn needs.

• Duplicate requests — Claude Code fires 2-5 identical background

requests every turn. Every one hits the API.

• Cache expiry — Anthropic drops the prompt cache after 5 minutes of

inactivity. Long pauses mean you start cold on the next turn.

Proposed Solution

cc-memory is a local Docker proxy that sits between Claude Code and
the Anthropic API. It intercepts every request and handles all of the
above before anything reaches Anthropic.

• Persistent memory — project context stored in ChromaDB and SQLite.

Relevant memory injected automatically on every new session. Claude
knows your stack, your decisions, your file structure when you come back.

• Smart compaction — adaptive threshold that adjusts based on content

density. Decision-heavy turns preserved longer. Recoverable tool
outputs evicted first. Compaction only fires when it actually makes sense.

• Tool token filtering — 55,000 tokens of tool definitions reduced to

approximately 2,400. Only what the current turn needs.

• Deduplication — duplicate background requests intercepted and replayed

from cache before they reach the API.

• Keepalive — pings Anthropic every 55 minutes to keep the prompt cache

warm so you never start cold mid-session.

• Context compression — long sessions summarized via free Cerebras calls

before hitting Anthropic. No token cost from your subscription.

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

MCP server integration

Use Case Example

A developer working on a multi-file project across multiple days.
Day 1: Claude learns the stack, the database schema, the conventions.
Day 2: new session, Claude already knows everything. No re-explaining.
No re-reading files. No contradicting yesterday's decisions.
The session starts exactly where the last one ended.

Additional Context

Built entirely using Claude Code over 60+ hours of development.
cc-memory was running throughout — Claude remembered every architectural
decision and every file across the entire project while building it.

Free, open source, runs fully local via Docker.
Your API key never leaves your machine.

Repo: https://github.com/AbdoKnbGit/claude-code-memory