Skip to content

Memory Portability Map

filhocf edited this page Sep 30, 2026 · 1 revision

Memory Portability Map

What this page is. A working map of who is observing what while we make memory portable across harnesses — not a statement of supported integrations. What we actually ship and support lives in docs/integrations.md and the READMEs. When a cell here turns into shipped code, it moves there through a normal PR and stops being a map entry.

How to read a cell. Every cell that says something exists links to it — a file on main, a PR, or an issue. Cells that describe a harness's own behaviour (e.g. "keeps writing to its native store after the MCP server is added") name the version observed and the date, because those drift with each harness release. A cell without a link or a date is a claim not yet backed — treat it as "not observed yet".

Discussion and decisions happen in #1364. Work items come out of this page one at a time.

The five layers

Reaching a store over MCP is not the same as a harness using it well. For a harness that ships its own memory, portability is a stack, not a plug:

  1. Connectivity — the harness can reach the store (MCP/REST).
  2. Connector / plugin — an installable that wires the store into the harness.
  3. Lifecycle hooks — session start / capture / injection / session end.
  4. Preference override — the harness prefers this store over its native memory (not just reachable alongside it).
  5. Usage toolkit — a skill/manual teaching the agent to use the store well: when to recall before acting, how to tag, when to save, rating.

Map (harnesses run daily by @filhocf)

Legend: ✅ shipped (linked) · 🟡 partial (linked) · 🔴 gap · — n/a

Harness 1. Connectivity 2. Connector/plugin 3. Lifecycle hooks 4. Preference override 5. Usage toolkit
Claude Code ✅ MCP ✅ claude-hooks/ ✅ claude-hooks/ (start, auto-capture, session-end) 🔴 hooks inject context but do not switch native file memory off (per maintainer, #1364, 2026-09-29) 🔴 no usage skill shipped; when-to-recall / how-to-tag live in each user's CLAUDE.md (per maintainer, #1364)
OpenCode ✅ MCP ✅ opencode/memory-plugin.js ✅ port of the Claude hooks (load, capture on message.part.updated, session.idle summary, harvest) 🔴 not observed 🔴 no usage skill shipped
Kiro CLI ✅ MCP 🔴 no connector shipped 🟡 harvest parses CLI legacy + v4 payload-wrapped (#1366); discovery of workspace layout in #1378; no session hooks 🔴 not observed 🔴 no usage skill shipped
Kiro IDE ✅ MCP 🔴 no connector shipped 🟡 IDE sessions migrated into the CLI workspace layout, read via the v4 parser (observed 2026-09-29, this host) — covered by the Kiro CLI row 🔴 not observed 🔴 no usage skill shipped
Kiro Crew ✅ MCP 🔴 no connector shipped 🔴 sessions live in data.sqlite3 (conversations_v2); SQLite reader in #1379, routing into automatic harvest is open (source-identity layer) 🔴 not observed 🔴 no usage skill shipped
OpenClaw ✅ REST / MCP 🟡 plugin @filhocf/openclaw-mcp-memory (v0.1.5 hit native better-sqlite3/sharp build issues; v0.2.0 planned as thin HTTP client, zero native deps — observed 2026-06) 🟡 openclaw-trajectory harvest incl. preamble filter; the active-memory post-response hook wrote empty assistantTexts (bug observed 2026-06, mistake-note) 🔴 not observed 🔴 no usage skill shipped
Hermes ✅ MCP (http://127.0.0.1:8000/mcp, observed 2026-09 on v10.73 client) 🔴 no connector shipped 🔴 no lifecycle adapter 🟡 ships a native SQLite memory; after adding our MCP it kept writing only to its own SQLite until the behaviour was changed to prefer ours (observed on a Hermes-based agent, 2026-09) 🔴 no usage skill shipped

Import side (bring your memory)

MemoryImporter (sync/importer.py, driven by scripts/sync/import_memories.py) already imports JSON exports with content-hash dedup and preserves created_at/updated_at. What is missing is a converter from each foreign schema into that format — N adapters + M converters, not N×M.

Source system Converter Status
mem0 mem0 export → MemoryImporter JSON 🔴 tracking issue is the next step (four RFC points), first converter
letta / zep / others — 🔴 demand-first (a real export from someone who runs it)

Related work

  • #1346 — harvest design-extraction (coverage of dropped reasoning); locale handled via existing config/locale.py mechanism.
  • #1345 — delta-sync (multi-writer authorship).
  • agent_id (#1100 / [#1278] / [#1297], merged) — per-memory authorship; the entry side (which agent/incarnation produced a session) is the source-identity layer feeding layer 2/4 here.

Seeded 2026-09-30 by @filhocf. Cells reflect one operator's observations; add your harness rows with links/dates.

Clone this wiki locally