mirror of
https://github.com/tinyhumansai/openhuman.git
synced 2026-07-29 14:02:19 +00:00
86 lines
3.3 KiB
Plaintext
86 lines
3.3 KiB
Plaintext
From: Sanil Jain <sanil@tinyhumans.ai>
|
|
To: Asha Mehta <asha@tinyhumans.ai>, Ravi Kulkarni <ravi@tinyhumans.ai>
|
|
Cc: OpenHuman Core <core@tinyhumans.ai>
|
|
Subject: Re: Memory integration plan for OpenHuman desktop
|
|
Date: Tue, 12 Mar 2026 09:14:00 +0530
|
|
Thread-Id: memory-integration-2026-03
|
|
|
|
Hi Asha and Ravi,
|
|
|
|
Quick summary after today's sync:
|
|
|
|
1. We should keep JSON-RPC as the transport for the desktop core.
|
|
2. The memory layer in the Rust core should use namespace as the main scope key.
|
|
3. We do not need user_id in the local storage contract for the current desktop runtime.
|
|
4. The frontend can adapt to richer result payloads as long as they still arrive inside JSON-RPC result.
|
|
|
|
Current work items:
|
|
- Ravi owns the Rust memory API alignment for list, delete, query, and recall.
|
|
- Asha owns the Neocortex v2 ingestion experiment using the GLiNER relex model.
|
|
- Sanil will review response models so they follow the Neocortex API style.
|
|
|
|
Important project facts:
|
|
- Project name: OpenHuman
|
|
- Subproject: memory-layer-completion
|
|
- Target milestone: March 22, 2026
|
|
- Preferred embedding model for local experiments: text-embedding-3-small
|
|
- Preferred extraction mode to try first: sentence
|
|
|
|
Known constraints:
|
|
- The desktop app is local-first.
|
|
- Core RPC currently binds to localhost only.
|
|
- We should avoid introducing user_id into every memory request unless we later support multi-user or remote runtimes.
|
|
|
|
Action items:
|
|
- Ravi: draft typed request/response structs for memory.query_namespace and memory.recall_namespace by Friday.
|
|
- Asha: prepare two ingestion fixtures, one Gmail-like and one Notion-like, with enough structure to test entity and relation extraction.
|
|
- Sanil: decide whether memory.init becomes a no-op compatibility method or is removed from the frontend wrappers.
|
|
|
|
One durable preference to remember:
|
|
I prefer keeping the memory core simple first and delaying graph traversal until after ingestion and recall are stable.
|
|
|
|
Thanks,
|
|
Sanil
|
|
|
|
---
|
|
|
|
From: Asha Mehta <asha@tinyhumans.ai>
|
|
To: Sanil Jain <sanil@tinyhumans.ai>, Ravi Kulkarni <ravi@tinyhumans.ai>
|
|
Subject: Re: Memory integration plan for OpenHuman desktop
|
|
Date: Tue, 12 Mar 2026 08:41:00 +0530
|
|
|
|
Agreed.
|
|
|
|
For the Neocortex donor path, I reviewed the neocortex_v2 extractor again:
|
|
- It uses a single GLiNER relex model.
|
|
- It supports sentence-level and chunk-level extraction.
|
|
- It adds recipient and spatial relation heuristics.
|
|
|
|
I think we should preserve those heuristics when we port the ingestion flow into OpenHuman.
|
|
|
|
Also, please record this:
|
|
- Ravi prefers narrower worker ownership to avoid merge conflicts.
|
|
- I prefer evaluation fixtures that include dates, owners, and product decisions.
|
|
|
|
Regards,
|
|
Asha
|
|
|
|
---
|
|
|
|
From: Ravi Kulkarni <ravi@tinyhumans.ai>
|
|
To: Sanil Jain <sanil@tinyhumans.ai>, Asha Mehta <asha@tinyhumans.ai>
|
|
Subject: Re: Memory integration plan for OpenHuman desktop
|
|
Date: Tue, 12 Mar 2026 08:09:00 +0530
|
|
|
|
One more note before I start:
|
|
|
|
- I will treat namespace as mandatory for memory query and recall.
|
|
- I will treat memory file APIs as optional until the core contract settles.
|
|
- I want the Gmail importer to preserve subject, sender, recipients, and sent_at metadata.
|
|
|
|
Dependency note:
|
|
- The frontend wrapper work depends on finalizing the result shape from the Rust core.
|
|
- The ingestion evaluation can run in parallel once the storage mapping is clear.
|
|
|
|
Ravi
|