Skip to content

Claude Code — retention & auto-save

How to make sure your work keeps flowing into HydraMem across Claude Code sessions — both backfilling existing knowledge and auto-saving going forward.

Honest scope

Unlike transcript-mining memory tools, HydraMem does not parse Claude Code's proprietary JSONL session files. It ingests Markdown and agent-submitted knowledge. So "retention" here means: keep your notes ingested, and let the agent remember decisions. This is deliberate (agent-driven, verified storage) — see ADR-0005.

1. Backfill existing knowledge

Point HydraMem at any directory of Markdown notes / docs and ingest it:

hydramem ingest ~/notes        --project myproject
hydramem ingest ./docs         --project myproject
hydramem ingest ./kms          --project myproject   # the scaffolded knowledge dir

Re-ingesting is safe — chunks are upserted by content hash (idempotent).

2. Auto-save with a Stop hook

Claude Code can run a shell command when the agent stops. Use it to re-ingest your knowledge directory so freshly written notes land in the palace automatically.

Create ~/.hydramem/hooks/save.sh:

#!/usr/bin/env bash
# Re-ingest the project's knowledge dir after each Claude Code turn.
# Honest: this ingests Markdown notes, it does not scrape the chat transcript.
set -euo pipefail
KMS_DIR="${HYDRAMEM_KMS_DIR:-$PWD/kms}"
PROJECT="${HYDRAMEM_PROJECT:-default}"
[ -d "$KMS_DIR" ] || exit 0
nohup hydramem ingest "$KMS_DIR" --project "$PROJECT" >/dev/null 2>&1 &
exit 0

chmod +x ~/.hydramem/hooks/save.sh

Wire it in .claude/settings.local.json (project) or ~/.claude/settings.json (user):

{
  "hooks": {
    "Stop": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "/absolute/path/to/.hydramem/hooks/save.sh",
        "timeout": 30
      }]
    }]
  }
}

Restart Claude Code (hooks load at session start).

The highest-signal memories come from the agent itself. Add a line to your project's CLAUDE.md / system prompt:

Before answering questions about past work, search HydraMem (hydra_search_tool). When we make a decision, call remember to persist it.

This captures verified decisions, not raw chat noise.

Verify

# After a session:
hydramem search "a phrase you discussed" --project myproject
hydramem stats --last-7d
tail -n 20 ~/.hydramem/metrics.db 2>/dev/null || hydramem stats --raw | head

Notes

  • hydramem ingest is idempotent — re-running is safe.
  • Keep private notes private; HydraMem stores everything locally by default.
  • Want true transcript capture? That's potential future work, tracked honestly — not a shipped feature today.

See also