plur-ai/plur/effective-memory

Effective Memory

Your agent has memory. These habits make it actually useful.

Without them, memory is a growing pile of assertions nobody retrieves. With them, memory compounds — each session builds on the last, corrections stick, and the agent gets measurably better over time.

These engrams cover the meta-rules every agent needs regardless of domain: how to capture corrections, when to recall before answering, what "verified" means, how to stay safe with destructive actions, and why never to type a weekday from memory.

They were written to be pinned — always eligible for injection, bypassing keyword gating. That is no longer what happens. Install strips pinned from every pack, this one included, so these engrams are matched on keywords like any other (injection_policy: on_match, with the match_terms above). The stripping is right — a pack author should not be able to occupy a recipient's context unconditionally — but it means this pack's meta-rules only surface when a session's wording happens to touch them.

Install

This pack ships inside @plur-ai/mcp. With both packages installed globally, preview it and install it by path (Bash or zsh):

plur packs preview "$(npm root -g)/@plur-ai/mcp/packs/effective-memory"
plur packs install "$(npm root -g)/@plur-ai/mcp/packs/effective-memory"

In PowerShell:

$packPath = Join-Path (npm root -g) '@plur-ai/mcp/packs/effective-memory'
plur packs preview $packPath
plur packs install $packPath

This needs @plur-ai/cli and @plur-ai/mcp installed globally (npm i -g @plur-ai/cli@latest @plur-ai/mcp@latest). plur packs install effective-memory, by name, does not work: packs install takes a folder or an archive, not a pack name.

plur init does not install this pack. plur-mcp init (the MCP package's own setup) does install it, and upgrades it when a newer version ships.

In the default lean MCP profile, preview/install use plur_admin actions plur_packs_preview and plur_packs_install, with { "source": "/path/to/pack" } as args. Use action help for the current schemas.

What's inside

12 engrams covering:

• Capture — call plur_learn immediately on corrections; detect correction-shaped phrases.
• Recall — plur_recall before factual answers; don't confabulate.
• Session lifecycle — bookend with plur_session_start / plur_session_end; plur_feedback on injected engrams; plur_timeline for long-horizon agents.
• Verification — artifact-first; never bulk-mark as done from narrative text.
• Safety — irreversible actions need actual user confirmation and one-item dry-runs.
• Discipline — read before edit; don't ask "want to continue?" mid-task.
• Time — never type a day-of-week from memory.

Why these were written pinned, and what happens instead

Pinned engrams bypass the keyword-relevance gate in scoreEngram and the per-pack and per-domain caps in fillTokenBudget. Cross-cutting meta-rules are the case that justifies it: "call plur_learn when corrected" is relevant to every session and keyword-matches almost none of them.

Install strips the flag, so that is not the behaviour you get. sanitizePackEngrams removes pinned and downgrades commitment: locked from every installed pack, and it is right to — a pack is an archive from a stranger, and letting its author decide what is permanently in front of your model is a privilege no producer should take by shipping a YAML field.

The consequence is that this pack's rules are keyword-gated, which is a weaker guarantee than they were designed for. Whether a host should be able to grant always-inject to a pack it trusts deliberately — as opposed to a pack claiming it — is an open question, tracked in plur-ai/plur#1019.

Versioning