LoCoMo benchmark — Claude Code memory paths

Four reproducible evaluation paths over the LoCoMo-10 dataset, each exercising a different way of giving Claude Code long-term memory.

All four use the same QA / judge / stats pipeline (eval.py → judge.py → stat_judge_result.py) so accuracy numbers are directly comparable.

Prerequisites

1. LoCoMo data — drop locomo10.json into .tmp/:
   Copy

   benchmark/locomo/claudecode/.tmp/locomo10.json
   

2. Python deps — managed by uv (use uv run invocations as below).
3. Claude Code CLI on PATH (the runners call claude -p).
4. An Anthropic-compatible model endpoint. Anything that speaks the Anthropic Messages API works. The original results were collected against doubao-seed-2-0 via Volces ARK:
   bashCopy

   export ANTHROPIC_AUTH_TOKEN=...
   export ANTHROPIC_BASE_URL=https://ark.cn-beijing.volces.com/api/compatible
   export ANTHROPIC_MODEL=doubao-seed-2-0-code-preview-260215
   

5. OpenViking server + plugin (only for SDK iso / SDK no-iso / e2e modes):
   • Start openviking-server (defaults to 127.0.0.1:1933). The e2e mode wants it running inside a tmux session so the runner can restart it across snapshots — set OPENVIKING_SERVER_TMUX=<session> (default ovserver).
   • Point at your clone of claude-code-memory-plugin:
     bashCopy

     export OPENVIKING_PLUGIN_DIR=$HOME/Dev/OpenViking/examples/claude-code-memory-plugin
     

     This is referenced from config/ov-hooks.json and from scripts/auto-capture.mjs.
   • Optional: export OPENVIKING_CLI_CONFIG_FILE=$HOME/.openviking/ovcli-local.conf to pin the local OV target and avoid leaking prod URLs from the user-shell's claude wrapper.

Running

Single-sample smoke test (any mode, any conv id):

./run_prompted.sh   conv-26
./run_sdk_iso.sh    conv-26
./run_sdk_noiso.sh  conv-26
./run_e2e.sh        conv-26

Full 10-conv run — omit the conv id:

./run_prompted.sh
./run_sdk_iso.sh
./run_sdk_noiso.sh
./run_e2e.sh

Each script ends with the result of stat_judge_result.py printed to stdout and saved to .tmp/result-<mode>/summary.txt. Per-question CSVs land in .tmp/result-<mode>/qa_results.csv.

What each Python entry point does

• ingest.py — feeds each LoCoMo session as a claude -p invocation; CC's vanilla auto-memory writes MEMORY.md inside the project dir.
• import_to_ov.py — uses the OpenViking Python SDK to push LoCoMo conversations directly into OV. Per-sample namespace by default; pass --no-user-agent-id for shared namespace.
• ingest_e2e.py — opens one claude -p per LoCoMo session in --input-format stream-json mode and streams the messages one at a time through stdin. The plugin's Stop hook auto-captures per turn; SessionEnd fires once per LoCoMo session (one commit per session). The benchmark's scripts/auto-capture.mjs shim extracts the LoCoMo per-message timestamp so OV event archive dates line up with conv chronology.
• eval.py — drives QA. Per-question claude -p; reads MEMORY.md for Prompted mode, talks to OV via auto-recall hook + MCP for the OV modes.
• judge.py / stat_judge_result.py — LLM-based judging and per-category accuracy roll-up.

Layout

benchmark/locomo/claudecode/
├── README.md
├── run_prompted.sh                # 4 entry points, one per published mode
├── run_sdk_iso.sh
├── run_sdk_noiso.sh
├── run_e2e.sh
├── ingest.py                      # Prompted ingest
├── ingest_e2e.py                  # e2e stream-json ingest
├── import_to_ov.py                # SDK pre-ingest (iso + no-iso)
├── eval.py
├── judge.py
├── stat_judge_result.py
├── config/
│   ├── ov-hooks.json              # hooks settings, references $OPENVIKING_PLUGIN_DIR
│   ├── ov-mcp.json                # MCP server (viking)
│   ├── ov-ingest.conf             # auto-capture on, auto-recall on
│   ├── ov-qa.conf                 # auto-capture off, auto-recall on
│   └── sys-prompts/               # optional system prompts (unused in default flows)
├── scripts/
│   └── auto-capture.mjs           # LoCoMo-aware Stop-hook (per-message created_at)
└── .tmp/                          # gitignored; data, snapshots, results
    ├── locomo10.json              # ← drop dataset here
    ├── legacy/                    # historical scripts/configs/reports kept aside
    └── result-<mode>/             # output of a run

Historical iterations of the experiment (r8 … r14b snapshot scripts, fine/listed/min ingest variants, intermediate reports) live under .tmp/legacy/ and are not part of the published numbers.