LocalHarness — the memory plugin
Memory
← the plugins /memory
On by default; it runs a small embedding model on your CPU, and turning it off leaves core’s memory slot empty.
over time
- each turn it adds what it knows to the prompt, within a budget core sets.
- on demand the agent searches by meaning, or pulls one fact in full.
- each session ends with a one-line summary in memory.
- for good facts outlive the session, in a database per agent and a file you can read.
how it learns
- new facts start as trusted as their writer has earned.
- repeats and updates a fact seen again gains weight; a new value replaces the old, which stays on file.
- consolidation while idle it replays what happened and strengthens the facts involved.
- one ranking facts reach the prompt ranked by need, truth and stakes, until the budget runs out.
- forgetting is opt-in archiving unused facts is off by default; nothing is deleted.
- checked writes each write is read back, and a failed one fails loudly.
Setup
Turn it on from the session, fill in its settings, and it checks they work; your coding agent can do the hardware part.
~/project — localharness start
shipped
qwen3.6-27b ~/project
Plugins: memory, autoresearch
❯ /plugins enable memory
Restarting with memory on — your conversation is kept.
✓ memory enabled in ~/.localharness/overrides.yaml
Download the embedding model now (about 1.2 GB)? [Y/n]: y
✓ memory-embedding: downloaded Qwen/Qwen3-Embedding-0.6B
Checking it now:
✓ memory-db: ~/.localharness/agents/orchestrator/memory.db (schema 10)
✓ memory-embedding: embedding model Qwen/Qwen3-Embedding-0.6B is in the local cache
In a session, /memory shows what it keeps.
Restarted with memory on. Your conversation continues.
memory: on in this session
❯ /memory
Memory is empty — nothing has been remembered yet.
in a session: /plugins enable memory · from a shell: localharness plugins enable memory
paste into your coding agent to set it up for your hardware
Set up LocalHarness memory on this machine. Install LocalHarness with its embeddings extra, keeping the extras I already use (a reinstall keeps only the extras it names). Then run `localharness plugins enable memory` and answer yes to downloading the embedding model (about 1.2 GB, into the Hugging Face cache). The model runs on the CPU, so no GPU setup is needed. You are done when `localharness doctor` shows memory-embedding passing.Surfaces
where you reach it
- /memory in a session: browse by tag, show, search, forget or promote.
- localharness memory from the shell: list, show, edit, remove, archive or restore.
- memory_search · memory_get · remember the tools the agent calls in a session.
- /plugins enable memory inside a running session: until it is set up, asks what it needs and checks it; then restarts with your conversation kept.
- memory.enabled false removes the memory section, the tools and /memory.