Memory in agents — short, long, working

Memory in agents isn't one thing.

- Working memory — current conversation. Lives in the messages array.
- Short-term memory — recent sessions, user preferences, active tasks. Often a small DB or cache.
- Long-term memory — facts, preferences, decisions that should persist forever. Vector store or structured DB.

Different storage, different lifecycles, different recall mechanisms. Mixing them is how 'amnesia bugs' happen.

Long conversations bust context limits. The fix isn't 'bigger context'. It's compression.

Every N turns:
1. Summarise old turns into a short paragraph.
2. Replace old turns with the summary.
3. Keep the most-recent turns verbatim.

Frameworks call this ConversationSummaryBufferMemory or similar. The agent retains the gist of long sessions without paying long-context tokens.

For really long sessions, keep the summary in long-term memory and rebuild context per session.

Keep last K turns verbatim. Older turns get summarised into a single 'summary' message at the front. Re-summarise when older + summary exceeds budget.

Simple, robust, framework-agnostic. Plugs into any LLM API.

Agents have a habit of writing 'remember user said password is X' into memory. Don't.

My rule:
- Long-term memory stores user preferences and decisions, not credentials
- Filter out PII / secrets before writing to vector store
- Tag memories with sensitivity; refuse to retrieve when context is wrong

Memory leaks across sessions feel benign until they aren't. Ship with redaction from day one.

Day 76 done.

- 3 kinds of memory
- Rolling summary survives token limits
- 20-line implementation
- Redact before persisting

Tomorrow (Day 77): the agent failure modes I've seen most. Then we close week 11.