An older fact stopped being used
Nothing was deleted. Recalled memories are rendered in a fixed group order and the block is then cut to a character budget — 700 on a live call. Because the order is fixed, the last groups are what the cut removes: facts about people, then what earlier work produced. A memory that used to appear is usually still active and now sits past the end of the slice.
What the symptom looks like#
A preference Connect honoured for weeks stops being honoured. A detail it used to mention on calls is absent, while the same detail still shows up in email replies. The memory viewer shows the row, active, with a sensible body and a recent enough date. Nobody edited or deleted anything.
The asymmetry between channels is the tell. Email has room; a live call does not, and the same memory bank produces two different briefs because one of them is being paid for in latency.
What it means#
Two limits are at work, and they bite in sequence. First, recall selects a small number of rows for this particular piece of work — six on a call — reserving all but two of those places for standing instructions, corrections and permission choices, ranked by importance and then by how well they fit the request. Two places are left for ordinary context.
Second, whatever was selected is rendered into a block grouped by kind, in a fixed order, and that block is sliced to the channel's budget. The slice keeps the beginning. So the question *what dropped* has a structural answer rather than a judgement behind it.
| Position | Group |
|---|---|
| 1 | Corrections the owner has made |
| 2 | Standing instructions |
| 3 | Permission and autonomy choices already made |
| 4 | Preferences — how things are liked done |
| 5 | Past approvals and rejections |
| 6 | What is known about the people involved |
| 7 | What earlier work produced |
The budgets, and why they exist#
| Block | Budget |
|---|---|
| What Connect remembers | 700 characters |
| Who you are talking to | 600 characters |
| What has happened before | 500 characters |
| Knowledge | 4 entries, 2,000 characters |
| A single memory line | 400 characters, whatever its length in the viewer |
These are not arbitrary. Instruction size is measured and recorded on every call: a 7,500-character brief produced a first token in 1.2–1.8 s, and 9,600 characters produced 2.3–3.4 s. A caller hears that difference as hesitation. The budget is the price of an agent that answers promptly, and spending it on a fourth restatement of a preference is a bad trade.
What Connect completed#
- Recalled the memories that apply at every tier the work touches — workspace, channel, endpoint and contact — narrowest first.
- Reserved places for standing instructions and corrections so the owner's own directions are not crowded out by routine notes.
- Marked every recalled memory as used, with a timestamp, which is visible in the viewer.
- Rendered them in group order and applied the channel's budget.
- Produced a reply or a call using what fitted.
What Connect did not complete#
- It did not decide the missing memory was unimportant, and it did not rank it out — the cut is positional, not evaluative.
- It did not delete, disable or supersede anything. The row is untouched.
- It did not tell anyone the block had been trimmed; there is no warning for a budget that was reached.
- It did not apply the same trim to email, which is why the same fact can be honoured on one channel and absent on another.
What you can do#
Shorten the memory. Say the fact in one line.
Result A short memory survives the slice and leaves room for others. This fixes more cases than anything else on this page.
If it is genuinely a direction rather than an observation, record it as a standing instruction or a correction.
Result Those groups are at the front of the block and have reserved places in recall, so the slice never reaches them.
Delete duplicates that say the same thing in different words.
Result Duplicates occupy separate places and separate characters while adding nothing.
Scope it to the contact or the channel it actually concerns.
Result A contact-tier memory competes only in work involving that person, which makes it far more likely to be selected at all.
What an administrator can do, and when to escalate#
- Check that the line's memory setting is full. Where it is not, the remembered block and the prior-history block are omitted entirely rather than trimmed — a different symptom with the same appearance.
- Prune the workspace tier. Workspace-level memories apply to every piece of work and are the ones most often stale.
- Count the standing instructions. They have reserved places, so a large number of them squeezes ordinary context out of recall before any character budget applies.
- Read the recorded instruction size on a slow call. It is the fastest evidence for whether the brief is genuinely oversized.
Escalate when a short, correction-kind, contact-scoped memory still never appears. That is not a budget symptom — it points at recall scoping or at the memory not being where it appears to be, and Connect ignored what I told it is the page for it. If the viewer itself is empty, start at the memory viewer shows nothing.
Questions#
Can the budget be raised?
It is configuration rather than a hard limit, and raising it is a real trade against the first-token time every caller experiences. Before spending latency, try spending editing: most briefs that hit the ceiling do so because of duplication and length rather than because the workspace genuinely has 700 characters of essential memory.
Does the same trimming apply to Knowledge?
Yes, with different numbers — four entries and 2,000 characters on a call. Near-identical Knowledge entries are collapsed before the block is built, but a fifth genuinely relevant entry simply does not travel. Retrieving Knowledge for an answer has the detail.
How do I know a memory was actually used?
Recalled memories are marked used with a timestamp, and the viewer shows it. A memory with no recent use, on a topic that keeps coming up, is either worded so it never matches or is being cut — and its length usually says which.