Reading…
Checks quiet when well · a word first, colour second
Cost · last 14 daysCost · 14 days weighted-token proxy · not dollars
Memory’s own footprint
Live — the tap cost and fan-out only; verdicts arrive after a session distills
Detail — sessions, decay candidates, query cost
Sessions —
Decay candidates
Query cost — traversal wall time
Every session ends.
What it learned shouldn’t.
Thalamus distills each coding-agent session into a graph of claims — decisions, problems, solutions — and serves them to the next session that needs them. This page shows how that loop runs, what it costs, and what we can and cannot yet claim about it.
Thalamus distills each coding-agent session into a graph of claims and serves them to the next session that needs them. Here is how that loop runs, what it costs, and what we can and cannot yet claim about it.
Memory is written after the fact, and judged after the factWritten after the fact,
judged after the fact
A session works
An agent in Claude Code, codex or Cursor does a task, calling memory tools as it goes.
It ends and is distilled
A model reads the transcript and writes claims: what was decided, what broke, what fixed it.
The graph keeps it
Claims, open threads and sources, each with provenance — which session, which scope.
The next session recalls
Relevant claims are served into context. The tap records every injection and its size.
Why the Status page shows cost now and quality later
The tap sees a recall the moment it happens — how many claims, how many tokens. Whether any of it helped can only be judged once the session is over and distilled, so a recall made this morning has no verdict until tonight. Pulse keeps those two clocks apart rather than inventing a per-turn score.
Attribution
Was each recalled claim reflected in what the agent then did? Used, or ignored.
A session works
An agent does a task, calling memory tools as it goes.
It ends and is distilled
A model reads the transcript and writes claims: decided, broke, fixed.
The graph keeps it
Claims, threads and sources, each with provenance.
The next session recalls
Claims are served into context; the tap records each injection and its size.
Attribution
Only after that session ends: was each recalled claim used, or ignored?
The tap sees a recall the moment it happens. Whether it helped can only be judged once the session is over, so Status shows cost now and quality later — never a per-turn score.
Not one memory
Each expert has its own scope. A session pinned to one reads and writes only there; other experts’ knowledge reaches it as cited, labelled context, or by asking — a consultation ticket. This page was designed that way: the designer scope asked eval-methodology what may honestly be shown.
Each expert keeps its own scope. Others’ knowledge arrives cited and labelled, or by a consultation ticket — this page was designed through one.
Did the memory help? Mostly we can’t say yet — and here is whyDid the memory help?
Mostly we can’t say yet
A judge checks whether a recalled claim shows up in what the agent did next. The obvious number is how often it says yes. The useful number is how often it says yes by accident.
+3.7 points — the part topic alone doesn’t explain+3.7 points is what the link adds;
the rest is topic overlap.
Most of the raw rate is the claim and the session being about the same thing.
The placebo keeps the topics and breaks the link, so the gap is what the link adds.The placebo keeps topics, breaks the link.
The same judge, uncorrected, calls 66% of injected tokens used (95% CI 59.5–72.8). The placebo says most of that is topic overlap.Uncorrected, the same judge says 66% (95% CI 59.5–72.8).
NOT YET VALIDATEDA waste rate means nothing without its sizeA waste rate means nothing
without its size
Two sessions. One wasted a quarter of what it recalled, the other half. By rate the first looks twice as good. By size the second wasted about 120 times as much. Pulse never shows a rate without the tokens beside it.The first looks twice as bad by rate and wasted ~120× as much by size. Pulse never shows a rate without its tokens.
Things the system has noticed about itselfThings the system has
noticed about itself
Is a pinned expert better than asking one? Not answerable yetPinned vs consulted expert?
Not answerable yet
The judge agrees far more often inside one project than across two:
A pinned session is judged within its scope; a consultation, across. Comparing the two would measure the judge, not the expert. Each side needs its own placebo first — until then every expert reads “not comparable yet”.Judge agreement differs 12× between the two; comparing them measures the judge.