Summarisation and what it loses
The summary is doing its job. Two hundred turns of conversation are represented in a few hundred words, the character still refers to things from weeks ago, and per-turn cost stayed flat.
Then a user asks about the thing they told it in the second week — the name, the date, the exact phrase they used — and it is gone, while a vague impression of the surrounding conversation remains. Nothing malfunctioned. That is what compression does, and it does it in a direction you can predict.
What gets destroyed first
Summarisation is lossy compression over text, and its losses are systematic rather than random.
Specifics go before gist. A summariser preserves what a passage was about and discards the particulars that made it that passage. Names, numbers, dates, quantities, exact wording. This is exactly backwards from what a relationship needs, because the particulars are what feel like being known.
Recency wins over significance. Something mentioned once, months ago, and mattering enormously will be compressed away in favour of the general shape of recent conversation. Summarisers weight by how much text there is, not by how much the text mattered.
Emotional weight flattens. A difficult thing a user said once becomes a clause. A long stretch of small talk becomes a comparable clause. Nothing in the mechanism knows the difference.
Contradiction resolves silently. If a user said one thing and later said the opposite, a summary tends to produce one coherent account rather than preserving both and the change between them. The fact that a position changed is often the most significant thing in the history, and it is the first casualty of being made coherent.
Structure disappears. Who said what, in what order, in response to what. A summary is prose about a conversation, and the conversational structure — the thing that made it a relationship rather than a set of facts — does not survive.
Generational decay is the real problem
The single most important property, and the one that gets missed at prototype scale: summaries built from summaries lose more each time.
If you summarise turns 1–50, then later summarise that summary together with turns 51–100, then again with 101–150, the earliest history has now been through three compressions. Every pass removes specifics from what the previous pass left, and there is nothing to restore from because the originals stopped being included.
The consequence is precisely inverted from what you want. The oldest history degrades fastest, and the oldest history is the foundation of a long relationship. A user’s first conversations — how they described themselves, why they started, what they were dealing with then — are the parts that get photocopied the most times.
Two mitigations, and neither is free.
Summarise from originals, not from summaries. Keep the raw history and rebuild the summary from source each time rather than iterating on the previous one. This eliminates generational decay and replaces it with a cost that grows with total history — you are re-reading everything periodically, which is a real bill and it grows forever.
Hold a separate durable store for what must not be lost. The facts that matter get extracted and stored structurally, outside the summary, where compression cannot reach them. This is why summarisation and extraction are so often paired: the summary carries texture and the store carries the things you cannot afford to lose. It only works to the extent your extraction caught them.
The turn
THE TURN — summarisation
· Condense old history, carry the summary
→ effectively unbounded reach at roughly
flat per-turn cost.
· What it destroys
→ specifics before gist. Names, numbers,
exact wording, and the fact that
anything changed.
· Summarising summaries
→ decays generationally. Oldest history
degrades fastest — the opposite of what
a relationship needs.
· Rebuilding from originals each time
→ no generational decay, and a periodic
cost that grows with total history
forever.
· The summary itself
→ PAID EVERY TURN. It sits in every input
from now on, and every word of it
competes with the character definition.
· Pairing with an extracted-fact store
→ protects what you thought to protect.
Nothing else.When the summary is rebuilt matters as much as how
Two operational details that decide more than the summariser’s quality.
Rebuilding mid-conversation produces a visible seam. Before the rebuild, the character has exact recent turns. After it, those turns are prose. If the rebuild happens between two messages in an active conversation, the second message is answered from a materially different state than the first, and users notice — it reads as a sudden change of attention. Rebuilding at conversation boundaries rather than in the middle of one costs nothing and removes the artifact.
A rebuild is a lossy write with no undo. If the summariser has a bad turn — truncates, misreads, drops a section — the previous summary is gone and the originals may not be retained. Keeping the raw history is what makes a bad summary recoverable, and the moment you discard originals to save storage, every compression becomes final. Storage is the cheapest thing in this entire stack; that trade is almost never worth taking.
Detection
Distance-to-failure by fact type. Log which kinds of fact fail: proper nouns and numbers failing while topics survive is the summariser’s signature. Whole topics failing is a different mechanism.
Age of failing facts. If failures skew heavily toward the oldest history, you have generational decay and you are summarising summaries.
Summary length over time. Watch it across a long relationship. A summary that stays a constant size while history grows tenfold is compressing at an increasing ratio, and the loss rate is rising with it — usually invisibly, because per-turn cost stays reassuringly flat.
Contradiction survival. Take conversations where a user changed their mind and check whether the change is represented. It usually is not, and it is usually the thing you would most want kept.
What this costs and what it doesn’t fix
The summary is in every input from the moment it exists, which is the recurring cost and the reason it is kept short — and every word of it is a word not spent on the character definition. Rebuilding from originals adds a growing periodic cost on top. Neither number is dramatic per turn, which is exactly why both get underestimated until the arithmetic is done properly.
And summarisation does not solve memory. It converts a hard boundary into a gradual, direction-specific loss, which is usually the better failure — but it is still a failure, it lands hardest on the history you care most about, and a system built on the assumption that the summary is a faithful record of the relationship is built on something that was never claiming to be one.