Hindsight vs Honcho

Hindsight's paper describes memory as "a structured substrate for reasoning", split into four networks: world facts, agent experiences, entity summaries, and evolving beliefs. Retain has an LLM turn input into typed facts with entities, time, and relationships. A background worker then deduplicates them and builds observations grounded in evidence, plus mental models: standing answers that get rewritten as the bank learns and can be read without an LLM call (README).

Recall runs four retrievers in parallel, fuses the results, reranks them with a cross-encoder, and fills a token budget. Reflect reasons across memories to answer questions a lookup can't. Each user, agent, or project gets its own bank, isolated by design.

On Hindsight Cloud you buy credit packs from $10 and pay per operation: $10.00 per million tokens retained, $0.75 per million tokens recalled, and $0.05 per reflect call. Those rates took effect July 6, 2026 (billing).

Honcho stores each message as it was sent, labelled with the peer who said it. A peer is any participant: a person or an agent. A background process called the deriver reasons over new messages and writes conclusions about each peer, including things that follow from what was said but were never stated outright. When a user corrects a fact, the deriver writes a new conclusion with a timestamp, and the old one stays for now. Dreaming runs later, once enough new conclusions have built up, deletes conclusions a newer fact has replaced, and merges ones that say the same thing.

Your agent reads memory two ways. session.context() returns the representation, a summary, and recent messages inside a token budget, in about 200 milliseconds, and those reads are unlimited. peer.chat() takes a question and reasons over the representation to answer it, at one of five levels from minimal to max, priced per query from $0.001 to $0.50.

Writes are asynchronous. A conclusion lands after the deriver runs, in time for the next turn but not the same one. Each peer can also keep its own view of every other peer it has observed. A support agent's model of a customer and the customer's model of that agent are separate objects.

Hindsight is MIT-licensed. If your company's open-source policy rules out copyleft, it clears review and Honcho's AGPL-3.0 code doesn't.

Hindsight is lighter to try locally. pip install hindsight-api runs with an embedded Postgres, and its Hermes integration can run a local daemon that stops when idle. Honcho's self-host runs Postgres and Redis on Docker. Hindsight also covers more deployment targets, including Oracle 23ai, a Helm chart, and enterprise SSO and audit logs on its cloud.