😤 I though we as a community had moved on from flawed retrieval only RAG benchmarks as indicators of good memory
Agent memory gets harder the longer a chat runs.
We wondered: how well does Observational Memory hold up at 100K, 1M and 10M tokens? So we ran it on BEAM, the long-term memory benchmark, with default settings.
Turns out it has the top published scores at 100K and 1M:
1
0
0
468
I have never been at an organization where the bottleneck is the speed of my linter and I can only imagine the eldritch horror of a codebase like that
0
0
1
198
Vercel coming after @flydotio
Your plain 𝙳𝚘𝚌𝚔𝚎𝚛𝚏𝚒𝚕𝚎 now deploys on Vercel.
Add a 𝙳𝚘𝚌𝚔𝚎𝚛𝚏𝚒𝚕𝚎 or 𝙲𝚘𝚗𝚝𝚊𝚒𝚗𝚎𝚛𝚏𝚒𝚕𝚎 at your root, then Vercel selects the Container preset, builds it, and routes traffic to the image.
vercel.com/changelog/depl…
1
0
0
126
In a post REA world is closed source really a moat? Any sufficiently motivated actor can figure out how your system is working and get a naive implementation working in an afternoon.
1
0
1
118
This kind of sentiment is exactly why @honchodotdev is open source. If a system is going to analyze your interactions to build a robust representation of you, those methods should be transparent and accessible.
How can you trust a memory store that's a black box
How can you trust a memory store that's a black box
my biggest frustrations with ai tools rn is their opaque memory systems.
chatgpt knows some things about me, claude remembers smth else, claude code has a different system, codex feels like a sprawling web of long-forgotten chats, and codebase docs + skills are yet another system.
i DO want these memories to be contextual and siloed. different models and harnesses have different strengths. i want their memories to reflect that.
but a few things that really bother me:
1. Opacity: how are memories created, curated, updated, and forgotten? i have almost no visibility into this.
2. Context rot: what random memories and old chats are seeping into the current context window?
3. Curation: no tools to curate memories. i'd like to have powerful tools to inspect, edit, prune, and shape these memories.
0
0
1
437
I've seen a lot of discourse talking about how your agent can just manage markdown files as a proxy for memory, but that's millions of tokens you are having a frontier model process which will massively balloon your costs
1
0
3
275
@honchodotdev let's achieve much stronger performance on recall benchmarks, let's you move past them to work on more predictive tasks (inferring intent, personalizing responses), while costing a fraction of a markdown system
0
0
1
47
Convinced agent swarms hacking all of the internet's critical infrastructure this week is just a PR stunt for the new deepseek model
0
0
1
101
Vineeth Voruganti retweeted
This is how I'm monitoring my Hermes agents:
- Set up @checkmk
- Configure a custom script to monitor:
- AI model subscription (OAuth)
- Cron jobs
- Memory health
- ...and whatever else I can think of.
I show how to do this in my latest video. (link in reply)
- Set up @checkmk
- Configure a custom script to monitor:
- AI model subscription (OAuth)
- Cron jobs
- Memory health
- ...and whatever else I can think of.
I show how to do this in my latest video. (link in reply)
17
23
250
17,151
Vineeth Voruganti retweeted
Two days to build an AI employee that watches several @orgo computers at once and stays on top of every decision, and Honcho is its memory 🫡
for everyone asking about the stack:
design: @paper
harness: @OpenAI
memory: honcho.dev
computers: @orgo
messaging & phone calls: @AgentPhoneHQ
connectors: @composio
email: @agentmail
GTM people & company search: @treg_ai@treg_aink that's everything
built everything in 2 days using @theo t3 and opus 5.5 high
1
5
21
959
Love to see more adoption of llm dreaming!!
Dreaming is going mainstream 🫡
Honcho has been dreaming since last fall: while your agents are idle it merges duplicates, retires facts that changed, and reasons out the patterns no single session shows.
Open source, and one memory across Claude Code, Codex, Cursor, Hermes and more: go.honcho.dev/ueHAdhR
0
0
2
188
Vineeth Voruganti retweeted
Honcho 3.2.2 is another pass on what gets remembered, so your conclusions read more like the notes you'd take yourself.
Tell your agent on a Monday "let's ship the beta Friday," and the conclusion carries the actual date. Say "sounds good" when it proposes a plan, and the plan Thanks" and "got it" stay conversation. Each conclusion holds one fact, with the details that make it useful attached.
Listing a peer's sessions got faster too: about 29ms down to 0.15ms on a 326k-row test table.
Self-hosting? v3.2.2 includes a migration, details here: go.honcho.dev/9LLGzQe
Tell your agent on a Monday "let's ship the beta Friday," and the conclusion carries the actual date. Say "sounds good" when it proposes a plan, and the plan Thanks" and "got it" stay conversation. Each conclusion holds one fact, with the details that make it useful attached.
Listing a peer's sessions got faster too: about 29ms down to 0.15ms on a 326k-row test table.
Self-hosting? v3.2.2 includes a migration, details here: go.honcho.dev/9LLGzQe
9
3
34
1,258
