Compare
Same problem. Two ways to handle memory.
Find bits of the past. Stuff them into the prompt. Hope it works.
1. Search
Hunt for whatever seems relevant.
2. Inject everything
Dump it all into the prompt.
3. Token cost
That dump often burns a lot of tokens.
4. No long-term reasoning
There’s no lasting picture to think with.
5. No sense of time
What changed, what’s outdated, and what still matters stay unclear.
Search. Paste. Hope.
Keep a living record. Reason over it when you ask.
Compact
Preferences, decisions, and open loops are kept, not dumped.
Reason
Archilas thinks with that record when it supports the answer.
Deliver
Clear answers land in the tools you already use.
Compact. Reason. Deliver.
Slack, GitHub, Gmail, Notion, Linear, Google Calendar, Google Drive, and AI chats are connected. Compatible with any AI that supports MCP. Cursor, Claude Code, ChatGPT, and Codex are examples.
Those systems keep memory an agent can look up. Archilas keeps the current decision and gives that context to the agent, with the source, the date, and the note a newer decision replaced. It abstains when the notes do not support the fact. How Archilas is different is the longer note. There is no head-to-head score on this page.
The preprint Evidence in Context Is Not Enough classifies individual failures of its own small-reader pipeline on one LoCoMo development split. In all 49 answer-stage failures, the gold evidence session was already in the prompt. 22 failures were retrieval-stage: 14 ranked too low, and 8 were dropped by the context budget. The paper does not claim that Archilas beats Mem0, or any other system.