Second Brain now creates its own memories.
Since our May 31 launch (v1.6), we shipped two major releases:
v1.7 added OAuth 2.0 on the MCP endpoint and importance-score-aware smart merging, so high-value memories resist being collapsed by near-duplicates.
v1.8 added semantic compression: a nightly cron job synthesizes entries with 10+ tags into net-new memory digests. Source entries are penalized in recall scoring, making the synthesized memory the authoritative version.
This is a meaningful architectural shift, not a UI update. Also shipped: full REST parity for recall/forget/list, and one-command client setup scripts.
66 views
Replies
At some point, more memories stop being useful. The challenge becomes figuring out which once are worth keeping around.
Second Brain for AI
@prince__kumar That’s the exact problem v1.8 is built around. The nightly compression job runs a tag sweep. Any tag with more than 10 entries triggers synthesis. The LLM writes one cohesive paragraph capturing current state for that area, stores it as the authoritative version, and the originals get a 0.4 score penalty in recall ranking so they stop competing with the digest. Fewer memories winning doesn’t mean losing anything. It means the right one surfaces.
Hello Rahil 👋 the semantic compression idea sounds really useful because AI memory can get messy quickly when too many small details pile up.
Second Brain for AI
@ali_haiider Yeah, the pile-up was the real problem in my own usage before I shipped this. I’d have 30 entries about the same project at different stages and recall would surface the oldest, most specific one. The compression job is basically what you’d do manually if you audited your notes once a week, except the LLM synthesizes them into a single current-state paragraph rather than just deleting the old ones.
Serand
How do you decide when a memory is important enough to keep separate versus merging into a larger summary.
Second Brain for AI
@anthony_adams_ Two things determine it. First, importance score: entries rated 4 or 5 out of 5 are excluded from compression entirely, so anything flagged as a critical decision or goal stays untouched. Second, the compression trigger is per-tag volume. Once more than 10 entries share a tag, that tag becomes a compression candidate on the nightly run. So a well-tagged project with lots of incremental updates gets synthesized, while a single high-importance decision entry just sits there accumulating recall history. The score is assigned at storage time by the LLM, or inferred if left blank.
The idea of memories creating their own better versions feels similar to how humans naturally keep important lesons and forget small details.
Second Brain for AI
@amard_sonal That’s the intuition I kept coming back to. Human memory doesn’t archive everything equally. It consolidates around what matters and lets peripheral details fade. Most AI memory tools just keep appending forever. The synthesis step is an attempt to build something closer to how retention actually works: the detailed originals don’t disappear, they just stop winning in recall.