Fixing Claude Code Context Compaction: Keep Memory Across Sessions

If Claude Code forgets your project decisions after compaction, that's not a bug — it's stateless context. Compaction summarizes the conversation and drops the details that matter: your conventions, constraints, and decisions. A persistent memory layer extracts those facts while you work and recalls them after compaction, so the next agent picks up right where the last one left off.

What compaction does — and why it forgets

Claude Code compacts long sessions by summarizing the conversation history into a shorter digest. That keeps the model running, but the summary is lossy: the exact wording of your decisions gets smoothed out, edge cases disappear, and "we decided to use bun, not npm" becomes a vague line about package managers — or nothing at all.

The deeper problem: compaction only preserves what was in the conversation. Anything the model never restated explicitly — your coding style, the architectural constraint you mentioned once, the library version you pinned — has no guarantee of surviving. Every developer who's re-explained the same convention after a compaction knows this pain.

Why a bigger window or /clear doesn't fix it

A bigger context window only delays compaction. Long enough sessions always compact — and the summary is just as lossy at 200K tokens as at 100K. Paying for more context to postpone the inevitable is expensive, and it doesn't change the failure mode.

/clear is worse: it wipes the conversation entirely, so you start from zero with no summary at all. Neither approach persists selected facts — that's a different job, and it needs a different tool. Background reading: RAG vs Memory Layer vs Vector DB.

The fix: a persistent memory layer

A memory layer sits outside the conversation. As you work, it extracts meaningful facts — preferences, decisions, constraints — and stores them as searchable memories. After compaction, your agent searches those memories with natural language and injects the relevant ones back into context. The summary can forget; the memory layer doesn't.

Concretely, this means the flow is:

1. You tell Claude Code: "we use bun, dark mode, and the API key lives in secrets.env".
2. The memory layer stores three facts.
3. Compaction happens. The conversation summary drops the details.
4. Your next message — or the plugin's automatic injection — searches memory for "project setup" and restores the facts into context.

5-minute setup: Audaxic + Claude Code

The fastest path is the Claude Code plugin: it auto-saves conversations and injects relevant memories into your prompts, so you don't have to remember to search.

npx @audaxic/memory-claude install

Then create a project and API key at memory-dashboard.audaxic.com (Hobby is free — 10K adds, 1K retrieve, unlimited projects) and follow the plugin prompts to connect your key. From then on, every session saves facts automatically.

Prefer MCP directly? Add the server to your Claude Code MCP config:

{
  "mcpServers": {
    "audaxic-memory": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://memory-mcp.audaxic.com/mcp",
               "--header", "Authorization: Bearer ak_your_key_here"]
    }
  }
}

Full guide: Claude plugin docs. Have older Claude sessions worth keeping? Run npm install -g @audaxic/memory-import to migrate past conversations with correct timestamps.

Before / after compaction

Before (no memory): compaction summarizes 3 hours of work. You ask "which package manager do we use?" — the agent guesses npm from training data. You correct it. Again. Every session.

After (with memory): compaction happens the same way, but the plugin injects "User uses bun, not npm" into the fresh context. You ask the same question — the agent answers correctly, citing your stored preference. Switch to Codex or OpenCode mid-project and the same facts follow you: switch from Claude Code to Codex to OpenCode — your agent picks up right where the last one left off. That's the whole point of one shared layer instead of per-tool memory.

FAQ

Written by Efe — Sep 03, 2026 · More articles · X @AudaxicTech