Giving OpenClaw Long-Term Memory — PowerMem 1.0.0 Officially Released

What Is PowerMem?

Large models are stateless; every conversation turn is a blank slate. The full-context approach is costly: inference slows down, token cost grows linearly, and the longer the context, the worse the model’s attention to the middle content (the lost-in-the-middle decay), so answer quality actually drops.

OceanBase PowerMem is an open-source intelligent memory system under the Apache 2.0 license, providing a persistent memory layer for LLMs and multi-agent applications. It distills key facts from conversations and persists them, automatically forgets stale ones, and recalls precisely when needed. Core capabilities:

• Hybrid retrieval: three-way recall across vector, full-text, and knowledge graph, so both semantic descriptions and exact keywords can hit.
• Ebbinghaus forgetting curve: models the human forgetting pattern, prioritizing recently used items and letting stale ones naturally fade.
• Intelligent memory extraction: the large model automatically distills facts from conversations, with deduplication, conflict updates, and related merging.
• Multi-agent support: independent memory space per Agent + cross-Agent shared collaboration.
• Multimodal: text, images, and audio can all be stored and retrieved.

We ran an evaluation on the LOCOMO benchmark (the standard dataset academia uses to measure AI’s long-conversation memory ability, simulating multi-turn long dialogues to test recall of historical information). PowerMem compared against the full-context approach:

Metric	PowerMem	Full Context	Improvement
Accuracy	78.70%	52.9%	+48.77%
p95 latency	1.44s	17.12s	91.83% lower
Token usage	~0.9K	~26K	96.53% saved

To validate it in a real scenario, we ran a comparison test on OpenClaw. By default, OpenClaw feeds the entire MEMORY.md into the system_prompt every turn, with no retrieval, and the content grows without bound as you use it. The PowerMem plugin replaces this mechanism, retrieving on demand before a session and intelligently extracting after it, putting only the relevant memory into context.

Experiment Group	Total Input Tokens
OpenClaw default (memory-core)	24,611,530
OpenClaw + LanceDB	51,574,530
OpenClaw + PowerMem plugin	4,533,508

For the same tasks, the PowerMem plugin’s token consumption is only 18% of the default approach.

---

Why Release 1.0.0

In v1.0.0, the API and integration methods are officially finalized, moving from “usable” into “production-ready.” This release delivers two layers of capability at once:

• The CLI (pmem) as the operations plane: a shared execution entry point for both humans and Agents, supporting orchestration and scripting.
• The Dashboard as the cognition plane: visualization, distribution analysis, and health monitoring of memory, turning data into a basis for judgment.

Agents need a low-friction operations plane to plug in, and humans need a cognition plane to understand the whole picture; only with both present is the product complete. We’ll expand on the thinking behind this layering in a dedicated article next week.

---

The OpenClaw Memory Plugin Is Now on ClawHub

The PowerMem long-term-memory plugin memory-powermem that we built for OpenClaw has been released. After installing it, OpenClaw gains cross-session long-term memory, retrieving relevant memory on demand before a session and injecting it into context, and intelligently extracting key facts to store after a session, no longer stuffing the entire MEMORY.md into the system_prompt every turn.

One-click install: install directly via the ClawHub Skill, and OpenClaw will automatically complete the plugin download, configuration, and slot switching:

https://clawhub.ai/Teingi/install-powermem-memory

Manual install: if you want to deploy it yourself, three steps:

1. Install and start the PowerMem service:

pip install powermem
# Start from a directory with .env configured
powermem-server --host 0.0.0.0 --port 8000

2. Install the plugin into OpenClaw:

openclaw plugins install memory-powermem

3. Modify the OpenClaw config (~/.openclaw/openclaw.json) to switch the memory slot to this plugin:

{
  "plugins": {
    "slots": { "memory": "memory-powermem" },
    "entries": {
      "memory-powermem": {
        "enabled": true,
        "config": {
          "baseUrl": "http://localhost:8000",
          "autoCapture": true,
          "autoRecall": true,
          "inferOnAdd": true
        }
      }
    }
  }
}

After restarting the OpenClaw Gateway, run openclaw ltm health to confirm connectivity.

A breakdown of the plugin’s principles and the full configuration guide will be published in a dedicated article next week.

---

Overview of What’s New in v1.0.0

The Operations Plane: CLI (pmem)

The CLI shares the same configuration (.env) and storage as the SDK and HTTP API; it’s the orchestration entry point for Agents and scripts.

pip install powermem  # or uv add powermem

# Memory operations
pmem memory add "User prefers dark mode" --user-id user123
pmem memory search "user preferences" --user-id user123
pmem memory list --user-id user123 -l 20

# Configuration management (interactive wizard, no need to hand-copy .env)
pmem config init

# Stats and ops
pmem stats --json
pmem manage backup -o backup.json
pmem manage cleanup --dry-run

# Interactive shell
pmem shell

The full command set covers memory (CRUD), config (view/validate/test/init), stats (statistics), manage (backup/restore/cleanup/migrate), and shell (interactive REPL), with bash/zsh/fish completion. See the CLI usage guide for details.

The Cognition Plane: Dashboard

A web visualization interface based on the same HTTP API, for viewing memory counts, user/Agent/type distributions, and system health status. v1.0.0 includes a number of fixes and experience improvements.

powermem-server --host 0.0.0.0 --port 8000
# Visit http://localhost:8000/dashboard/ in a browser

---

Multiple Integration Methods

v1.0.0 offers five integration methods, all sharing the same configuration and storage:

Python SDK (three lines to start):

from powermem import Memory, auto_config

config = auto_config()
memory = Memory(config=config)

memory.add("The user likes drinking coffee", user_id="user123")
results = memory.search("user preferences", user_id="user123")

CLI: operate directly from the terminal, the top choice for scripting and Agent orchestration.

HTTP API: RESTful + Swagger docs + API Key authentication, for any language.

MCP Server: supports the Model Context Protocol, so MCP clients like Claude Desktop can read and write memory directly.

Dashboard: web visualization, memory statistics, and analysis.

---

Try It Now

pip install -U powermem
# or
uv add powermem

pmem --version

• GitHub: github.com/oceanbase/powermem
• PyPI: pypi.org/project/powermem
• Discord: Join the community
• Issues / Discussions: Feedback and discussion

---

OceanBase PowerMem Team
2026.3