Repository navigation
Replies: 7 comments
|
Reference matrix to anchor the arc — a starting point, not exhaustive (market data is mid-2026, worth re-checking). "Use in any agent" side — harnesses and how they do memory today:
"Bring your memory" side — importer candidates (systems a neighbouring project already imports from): mem0, letta (ex-MemGPT), zep, honcho, hindsight, cognee, supermemory, plus an LLM-guided generic importer. Each cell (harness × memory-system) is a potential pair. Proposed first: Claude Code ⊕ mem0 — adapter exists, importer doesn't. Which cells matter most to you and the community is exactly what I'd like this discussion to surface. |
|
Before the scope question, I checked the matrix against Corrections to the harness side
So "only Claude Code has this today" should read "Claude Code and OpenCode". I would also split the column in two, as above. Nearly every harness on the list can already reach the store over MCP. What most of them lack is the lifecycle part: capture, injection and session end. Folding both into a single "no" makes the gap look larger than it is, and it hides which work is actually missing. Corrections to the import side Importing is not absent. It is absent for foreign systems. On the per-pair model I agree with the two halves. I don't agree that the pairs are the unit of work. An importer knows nothing about the harness, and an adapter knows nothing about where a memory came from. "Claude Code + mem0" is the mem0 converter plus the existing Scope, concretely
A converter on existing import code is something we can take on now. A portability framework across N harnesses is not, at least not yet. Next step: if the converter scope works for you, open a tracking issue for the mem0 converter with the four usual RFC points (problem, alternatives such as a generic interchange format instead of per-system converters, compat impact, first-PR scope), and I will review against that. |
|
Thanks for checking the matrix against The matrix was a starting proposal, not a claim of scope. Its job was to open the question "which pairs matter?", not to declare what exists. So I take the factual corrections as-is (OpenCode has a lifecycle plugin; Kiro/OpenClaw harvest is real; But "reachable over MCP" is not "prefers this store", and that gap is bigger than one So the harness side isn't two states (MCP-access, hooks). For a harness with native memory it's a stack: (1) connectivity → (2) connector/plugin to install → (3) lifecycle hooks → (4) preference override (native → ours) → (5) a usage toolkit (skill/manual) that teaches the agent to use the MCP correctly — when to recall before acting, how to tag, when to save, rating. Layer 5 is the one a plain Everyone works within their own reach. I run Kiro, Hermes, OpenClaw, OpenCode and DeepSeek daily, so those are the harnesses I can actually observe, find memory patterns in, and figure out how to import from — which matches your own rule (an adapter comes with someone who uses that harness). I'm not proposing eight speculative converters. The real proposal: a living page instead of a comment thread. A discussion is good for debate but bad for accumulated state — comments don't edit well, the matrix can't evolve in place, and per-harness progress (native memory? connector? hooks? skill? preference override? import path?) has nowhere to live. I'd like a wiki page as the source of truth we edit as patterns emerge, with the discussion staying the place we argue and decide. Issues and PRs then come out of the wiki one at a time (the mem0 converter being one of them), so the broad map doesn't inflate the work breakdown. Where should it live? The repo wiki is enabled. Tell me the path you'd want this under — a wiki page, a |
|
The five layers are a better model than my two columns, and I'll take them. Layer 4 also applies to Claude Code. Claude Code keeps its own file-based memory next to ours, and On where it lives: the wiki. You have write access to the repo, which covers the wiki, so you can edit it directly without a PR round-trip for every cell. Please call the page
The discussion stays the place where we argue and decide, as you said, and work items come out of the page one at a time. The mem0 converter is still the first one. The tracking issue with the four RFC points is the next step there, independent of the page. Next step: you seed |
|
Following up with numbers, since the harness side of portability is easier to argue with data than in the abstract. I extended the Phase 0 coverage instrument (#1350) to also record detected language per text block (cheap pt/en heuristic, measurement not inference), and taught discovery + parsing to reach every Kiro session format on this host. Here is what a single harness — Kiro — actually looks like. Three session formats, one harness:
The old IDE Coverage + language, measured on the real archives (fork, feeding the #1350 instrument):
Two things the numbers say:
None of this is a portability framework. It is one harness, three formats, and the plain work of reaching and reading each — before any adapter or preference-override even enters the picture. That is the point I was trying to make earlier, now with a table instead of an assertion: "reachable over MCP" is a long way from "harvested well", and language is a dimension the English-first tooling doesn't see. The discovery gap and the unread SQLite source are bugs, not features — the parser could already read that content, harvest just never reached it. Both fixes are on my fork; I'll open a tracking issue and a PR for each (smallest first), one at a time. Posting the numbers here first, as you prefer. |
|
Both gaps are real, I checked them against Two things about where this goes, though. The numbers belong in #1346 and in the two issues, not here. They are harvest coverage for one harness. This thread is about portability, and what we agreed here two comments ago is still open: the On the language column: these are your own archives, so the 85-89% pt-BR is you, not the user base. It shows that the extractor has to handle the locale of the person running it. It does not show that most of the content anyone harvests is Portuguese. We already made that call for everything else: harvest patterns, NER and NLI patterns load per locale ( Next step: #1376 as the first PR. Here, a yes or no from you on the wiki page and the mem0 issue. |
|
Yes to both, and thank you — for the patience and for keeping the thread pointed straight. This arc crosses a lot of fronts at once (harvest coverage, language, the parsers, portability) and in the middle of that it's easy to drift; your corrections kept it honest. Done:
On the numbers: you're right, they belong in #1346 and the two issues, not here — moved my focus there. And the language correction is a real fix to my framing: the 85-89% pt-BR is my archives, so it shows the extractor must honor the operator's locale, not that most content is Portuguese. I've reframed the design-extraction RFC (v0.4) to require the extractor to follow the existing locale mechanism ( #1376 first, as agreed — it's green and ready. #1377 behind it. Work items come off the wiki one at a time from here. |
Uh oh!
There was an error while loading. Please reload this page.
Opening this as a design discussion before any code, per the usual flow. Full draft lives in my fork at
docs/rfc/rfc-memory-portability.md.The thesis. The service is a great memory destination today, but a real adoption journey has two ends we don't treat as one coherent whole:
store_memory/harvest.claude-hooks/).The key point: these two only deliver value reconciled per pair (harness × memory-system). Concrete case: "someone uses Claude Code + mem0" needs two chained actions — (a) migrate mem0's memory into the store, and (b) integrate the store as Claude Code's native memory. Either one alone is half a solution.
Why this belongs here, not feature creep. The foundation is already merged — agent_id (#1278/#1297) gives portable identity. And a neighbouring project already does exactly this arc: it ships importers for ~8 memory systems and native per-harness adapters. We don't. That's the strategic adoption gap: broad import/adapter support is what lets someone try the project without losing what they already have.
Method I'd propose (before code): an N harnesses × M memory-systems matrix. For each prioritized pair, two characterizations precede code:
claude-hooks/.A feature-parity study across competing memory systems is part of this — research that precedes code, and honestly the part that would raise the project's bar the most.
First pair: Claude Code ⊕ mem0 (most popular harness + most popular memory layer; the adapter exists, the importer doesn't).
Ask: this is large and needs your read on scope before I code anything. Does the two-halves model + the per-pair method land as the right groundwork? Should the parity study be a separate track? I'll drop a reference matrix (harnesses × memory-systems) as a follow-up comment to anchor the conversation.
All reactions