Human memory is not a database. It is a lossy, reconstructive, emotionally weighted system that prioritizes differently depending on context. The best moments of engineering Tessera have been the moments where I stopped trying to build a perfect recall system and started trying to build one that remembers the way I do.
Not everything in the corpus is equally important. An email confirming a meeting time has near-zero long-term value. A decision to change a client’s backup architecture has value that persists for years. A reflection on why a particular approach to incident response failed has value that compounds over time as it informs future decisions.
The Salience Model
Tessera assigns a salience score to every artifact and every extracted decision. Salience is a function of four factors: consequence (how much did this affect outcomes?), recurrence (how often does this type of decision come up?), recency (how recently was this relevant?), and emotional weight (was this associated with a strong positive or negative outcome?).
High-salience artifacts are retrieved preferentially. Low-salience artifacts are deprioritized but not deleted. The salience scores decay over time unless reinforced by new references or similar decisions. An artifact that was highly salient two years ago but has never been referenced since gradually fades in retrieval priority.
This mimics something real about how expertise works. The remediation patterns I use most frequently are the most accessible in my own memory. The ones I used once, ten years ago, are still there but require more effort to recall. Tessera should work the same way.
Forgetting as a Feature
Controlled forgetting is essential for a life assistant. I do not want Tessera to treat a grocery list from 2019 with the same priority as a strategic decision from last quarter. The forgetting mechanism is not deletion. It is deprioritization through salience decay.
But some things should never decay. Decisions that establish precedent, patterns that define my operating philosophy, and reflections that capture hard-won lessons maintain their salience regardless of recency. These are tagged as foundational during enrichment, and foundational artifacts are exempt from decay.
The distinction between transient and foundational is itself a judgment call, and Tessera learns it from my behavior. When I reference an old decision in a new context, Tessera increases its salience and may reclassify it as foundational. When I explicitly override a past decision, Tessera notes the override and adjusts the pattern accordingly.
Working Memory
Beyond long-term memory, Tessera maintains a working memory: the active context of what I am currently doing, what I have recently discussed, and what decisions are pending. Working memory is ephemeral by design. It lasts for the duration of a task or conversation and then dissolves back into the long-term store.
The working memory is where the agentic layer operates. When Tessera is executing a retrieval plan for a remediation briefing, the intermediate results live in working memory. The plan itself, the partial results, and the reasoning trace are all held in working memory until the task completes. After completion, the final output is indexed and the intermediate state is discarded.
This is a deliberate design choice. I do not want Tessera to remember every query I ran during a stressful remediation. I want it to remember what I decided and why. The journey is noise. The destination is signal.