Activation & decay

Every engram has a number called activation. Activation decides whether the engram gets injected into the agent’s context for a given task. The model is borrowed from ACT-R (Adaptive Control of Thought – Rational), a cognitive architecture used in computational psychology.

The shape

A(t) = ln( Σ tᵢ^(-d) )

  tᵢ = time since the i-th access (in days)
  d  = decay rate (default 0.5)

Each time the engram is accessed (recalled, injected, or rated positively) its history gets a new timestamp. The activation is a power-law sum over the time-since-access for every past access. Fresh accesses dominate; old accesses still contribute, just less.

The practical effect: engrams that match real tasks repeatedly stay sharp. Engrams that surface but never help fade. Engrams you wrote two years ago and never reused decay so far that they essentially disappear from injection — without you ever having to delete them.

What strengthens activation

• Hybrid recall hit that returns the engram (plur_recall_hybrid)
• Injection use when the agent visibly relies on the engram (downstream signal: positive feedback or no correction)
• Positive feedback via plur_feedback ENG-... positive
• Explicit pin (pinned: true) — bypasses decay; injection regardless of relevance

What accelerates decay

• Time since last access
• Negative feedback via plur_feedback ENG-... negative
• Supersession — when a newer engram covers the same ground

Status transitions driven by activation

Activation range	Status drift
High, fresh accesses	active (default)
Below median, no recent access	candidate for dormant (next maintenance pass)
Very low + superseded	candidate for retired

You can run the maintenance pass manually:

plur batch-decay --commit

…or let it run on a schedule (Enterprise: nightly).

Why this matters

Most “agent memory” systems treat the store as a passive log: anything you saved is forever, and the only way to manage relevance is to delete things. That model breaks at scale — after a few months your store is full of stale rules and you can’t tell which still matter.

PLUR’s model lets the use of the system tell you what matters. You don’t curate. You correct, you give feedback, and the ranking sorts itself.

The feedback loop

Agent ranks engrams       →  injects top-N for the task
You react to the agent    →  positive/negative implicit signal
You give explicit rating  →  plur_feedback                  →  activation update
Next session              →  ranking has shifted

Three sources of signal: implicit (you didn’t push back), explicit (plur_feedback), and behavioural (the agent corrected itself after seeing the engram). Each is weighted differently in the activation calculation.

Next: Knowledge Packs →