Agent memory you can audit, replay and erase
What sets Delentia's memory apart is not accuracy or compression. It is a history you can check, revocation with a reason, and erasure of one person that can be proven. This page lists what was measured and what is not claimed.
Built and measured
- Memory is an append-only event log, hash-chained per person and written in the same transaction as the memory itself. Replaying the log gives back exactly the stored memories (checked from 100 to 5,000 operations), and any edited event is named.
- The head of the log is written into the audit chain, so rewriting the whole log is detected. You can look at a memory's history and see the store as it was at an earlier point.
- Erasure of one person: their memory text is sealed under a key held outside the database, and erasing means destroying that key after an approver signs. After that, scanning the raw database file finds none of their text, while other people's memories and the chain still verify.
- Revocation instead of silent edits: a memory is retired with a written reason and stays visible in the history. A person can pin a few standing rules that are shown to every episode, labelled as their own words and never counted as permission.
- Recall by meaning is opt-in (it needs a local embedding model). On a held-out set of 40 paraphrased questions it found the right memory in the top three for 37 (English, nomic-embed-text), against 6 for plain word matching; for Thai, bge-m3 reached 38 of 40.
Compared on one workload
- One synthetic workload of 2,000 operations, against LangGraph (SQLite checkpointer) and mem0: the log plus sealed store was 1.5 MB, LangGraph 192 MB, mem0 7.1 MB. All three could return an earlier state exactly. Only Delentia detected an edited record. After each system's own delete call, the person's text was still in the file for LangGraph and mem0; only Delentia's erase removed it.
Not claimed / not done
- No claim that Delentia memory is more accurate than other systems. On the same 30 questions mem0 found 24, Delentia's default word matching 7, and Delentia with embeddings 22.
- No compression claim. The earlier 74% figure was a formula. Measured in bytes the delta saves 39% to 90%, and ordinary compression of snapshots does about as well.
- No recall-latency claim. With embeddings on, recall costs a few hundred milliseconds.
- Erasure does not yet cover the separate skills database or audit rows written before sealing was added. Not tested with real users' memories. With about a thousand near-identical memories, ranking among look-alikes gets worse.
- No compliance claim. Software cannot make an organisation compliant with the PDPA; this only gives you tools to erase and to show what happened.