Summary

This PR adds a purpose-specific memory role-slot contract so memory plugins can compose instead of replacing each other through one global plugins.slots.memory selector.

What changes:

• Adds canonical memory role slots: memory.recall, memory.compaction, memory.capture, memory.dreaming, and memory.userModel.
• Keeps legacy plugins.slots.memory as a backward-compatible recall shorthand during an extended migration window.
• Adds per-agent memory role-slot overrides under agents.list[*].plugins.slots, limited to memory roles plus legacy memory.
• Wires role-slot selection through config schema/types/help/docs, plugin activation, gateway startup preloading, runtime memory resolution, skill publishing, validation, doctor repair/preview, uninstall cleanup, status diagnostics, and SDK-facing aliases.
• Preserves the existing global plugins.slots.contextEngine path for LosslessClaw and other context-engine plugins.
• Explicitly rejects unsupported per-agent agents.list[*].plugins.slots.contextEngine overrides, because agent-scoped context-engine semantics are out of scope for this PR.

Why this matters:

• Advanced OpenClaw users are already operating layered memory stacks: factual recall, context compaction, auto-capture, dreaming/consolidation, and inferred user modeling are distinct concerns.
• The previous single memory slot forced useful plugins into workarounds such as context-engine ownership, sidecars, or global replacement of the primary recall plugin.
• This branch makes the config contract match the memory architecture users are trying to run while keeping the migration path conservative.

Intended outcome:

• memory-core can remain the factual recall owner while other plugins own capture, compaction, dreaming, or user-model roles.
• Existing configs using plugins.slots.memory continue to work.
• openclaw doctor --fix copies non-conflicting legacy plugins.slots.memory values into plugins.slots["memory.recall"] but keeps the legacy key in place for plugin compatibility.
• Plugin lifecycle and diagnostics treat role-slot-selected memory plugins as intentionally active across startup, validation, skill publishing, runtime resolution, and cleanup.

Out of scope:

• This does not replace or remove plugins.slots.contextEngine; LosslessClaw and other context engines continue to use the existing global contextEngine slot.
• This does not add agent-scoped contextEngine semantics.
• This does not implement the owner-provenance object shape proposed in RFC: machine-readable memory-slot ownership — plugins.slots.<slot>.owner #70823.
• This does not require third-party plugins to implement every memory role immediately.

Linked context

Closes #60572.

Directly related memory-composition requests and prior proposals:

• Feature Request: Support for supplementary/additive memory plugins alongside the primary memory slot #38874: supplementary/additive memory plugins alongside the primary memory slot.
• feat(plugins): allow stackable memory plugins #2306: earlier closed PR proposing stackable memory plugins.
• [Feature]: More memory plugins #2080: earlier request for multiple simultaneous memory plugins and emerging memory architectures.
• Feature Request: Pluggable persistent memory backend for agent context #77714: pluggable persistent memory backend demand.

Dreaming and memory-role separation context:

• [Feature] Decouple dreaming from memory-core — allow any memory plugin to implement dreaming protocol #85473: future protocol work for making dreaming implementable by non-memory-core memory plugins. This PR adds the memory.dreaming slot contract but does not implement that full protocol.
• Allow memory-core dreaming runtime to coexist with third-party memory slot plugins #63067 and fix(plugins): exempt dreaming engine from memory slot fast-path in loader #65411: prior work around keeping dreaming behavior available when another plugin owns the legacy memory slot.
• fix(memory): respect selected slot in dreaming config #62275: previous selected-slot dreaming config behavior.

Runtime/diagnostic slot handling context:

• Memory runtime checks should load the configured memory slot #76567 and fix(memory): load configured memory slot for runtime checks #76495: configured memory-slot runtime checks should load/consult the selected memory plugin.
• [Bug]: openclaw doctor reports 'No active memory plugin is registered for the current config.' when memory-lancedb configured #78540: doctor/status memory diagnostics could misreport configured memory providers.
• Bug: doctor --fix does not clean up plugins.slots.* when plugin is missing, leaving config unrecoverable #76550: stale plugins.slots.* references could leave config unrecoverable until doctor cleanup was broadened.

Context-engine compatibility context:

• [Bug]: lossless-claw selected and enabled but not registered as context engine in 2026.5.2 #77063, [Bug]: lossless-claw context engine not registered in 2026.5.2 (Windows npm global, duplicate plugin id observed) #77185, [Bug]: Bug Report: Context engine plugin not registered despite being loaded #76576, and Bundled legacy context engine not registered in CLI compaction lifecycle worker — dreaming diary subagent fails silently (follow-up to #76576) #81843: context-engine registration and LosslessClaw-adjacent failures shaped the compatibility review for preserving the existing global plugins.slots.contextEngine behavior while rejecting unsupported per-agent contextEngine config.

Future API/RFC context:

• RFC: machine-readable memory-slot ownership — plugins.slots.<slot>.owner #70823: future machine-readable slot ownership/provenance. This PR intentionally keeps scalar plugin IDs and solves the memory role composition layer first.

Real behavior proof

• Behavior or issue addressed: Current-head proof refresh for PR feat(memory): add multi-slot memory role architecture #86210 at 23a29722524c982f9ed410610395e1db0363857e, covering the multi-slot memory role contract, the latest recall-default cleanup, Hermes migration regression coverage for plugins.slots["memory.recall"], legacy plugins.slots.memory doctor compatibility, per-agent memory role routing coverage, context-engine slot compatibility, and exact-head CI signal.
• Real environment tested: Exact-head deterministic/upgrade proof ran on the available paired proof-host route via [REDACTED] host [REDACTED], checkout [REDACTED], Node v22.22.2, pnpm 11.2.2, OpenClaw 2026.5.28 (23a2972), artifact [REDACTED]/proof-refresh-23a2972252-20260528T172630Z.
• Exact steps or command run after this patch: Fetched kklouzal:feat/multi-slot-memory-architecture, detached the proof checkout to 23a29722524c982f9ed410610395e1db0363857e, then ran corepack pnpm install --frozen-lockfile, corepack pnpm build:strict-smoke, focused memory/config/Hermes regression suites, agentic shard, plugin memory tests, command/config memory tests, extension memory tests, corepack pnpm tsgo:prod, git diff --check origin/main...HEAD, a legacy-memory doctor --fix upgrade harness, a plugins.slots.contextEngine = "lossless-claw" config-compatibility harness, openclaw models status --probe --json, GitHub Copilot live-auth probes, and Honcho availability probes.
• Evidence after fix: Redacted runtime logs are preserved on the proof host at the artifact path above; copied redacted excerpts below are included for maintainer-visible review. GitHub proof gate for the pushed head also passed: https://github.com/openclaw/openclaw/actions/runs/26588940267/job/78342338941.

[REDACTED] environment.txt
head=23a29722524c982f9ed410610395e1db0363857e
node=v22.22.2
pnpm=11.2.2
openclaw=OpenClaw 2026.5.28 (23a2972)

[REDACTED] redacted/03-focused-memory-tests.log
node scripts/run-vitest.mjs run extensions/migrate-hermes/provider.test.ts extensions/memory-wiki/src/bridge.test.ts extensions/active-memory/index.test.ts src/config/config-misc.test.ts src/config/config.plugin-validation.test.ts src/commands/doctor/shared/legacy-memory-slot.test.ts src/commands/status.scan.shared.test.ts src/plugins/effective-plugin-ids.test.ts src/plugins/loader.test.ts src/plugins/config-state.test.ts
Test Files 10 passed (10)
Tests 496 passed (496)

[REDACTED] redacted/04-agentic-shard.log
[test] passed 3 Vitest shards in 65.22s
Test Files 76 passed (76)
Tests 1260 passed (1260)

[REDACTED] redacted/05-plugin-memory-tests.log
Test Files 5 passed (5)
Tests 320 passed (320)

[REDACTED] redacted/06-command-config-memory-tests.log
Test Files 1 passed (1)
Tests 14 passed (14)

[REDACTED] redacted/07-extension-memory-tests.log
Test Files 1 passed (1)
Tests 126 passed (126)
Test Files 1 passed (1)
Tests 11 passed (11)

[REDACTED] redacted/10b-legacy-memory-doctor-compat.log
plugins.slots.memory: copied legacy memory slot to memory.recall (memory-core) and kept the legacy key for plugin compatibility.
plugins.slots.memory: legacy memory slot remains for plugin compatibility; use memory.recall as the canonical selector.
{"valid":true,"path":"[REDACTED]"}
{
  "legacy": "memory-core",
  "recall": "memory-core",
  "slots": {
    "memory": "memory-core",
    "memory.recall": "memory-core"
  }
}

[REDACTED] redacted/11-context-engine-lossless-config.log
{"valid":true,"path":"[REDACTED]"}
{
  "slot": "lossless-claw",
  "present": false,
  "status": null,
  "enabled": null
}

[REDACTED] command-summary.json
"head": "23a29722524c982f9ed410610395e1db0363857e"
"deterministicOk": true
"deterministicFailures": []
required gates included install, build:strict-smoke, focused-memory-tests, agentic-shard, plugin-memory-tests, command-config-memory-tests, extension-memory-tests, tsgo:prod, diff-check, legacy-memory-doctor-compat, context-engine-lossless-config, and models-status-probe.

• Observed result after fix: The exact current head validates the latest canonical recall-default behavior and Hermes migration regression, with extensions/migrate-hermes/provider.test.ts included in the 10-file focused suite. The upgrade harness confirms doctor --fix copies plugins.slots.memory = "memory-core" into plugins.slots["memory.recall"] = "memory-core" while retaining the legacy key for third-party compatibility. The role-slot deterministic gates, agentic shard, type gate, and diff check all passed at the current pushed head.
• What was not tested: Authenticated GitHub Copilot inference, Honcho ask/status, and installed LosslessClaw runtime execution were not available on this proof host at refresh time. The exact-head logs preserve the negative checks: github-copilot/gpt-5.4-mini direct inference failed with No API key found for provider "github-copilot"; the live-model harness reported no API keys found; skipping; lossless-claw config validation passed but the plugin was not installed in the derived registry; local Honcho health did not respond. The stale bb771d live output has therefore been removed rather than reused.
• Proof limitations or environment constraints: The host route available through this lane exposed the dedicated PR checkout and deterministic proof environment, but not the expected live github-copilot, Honcho, or LosslessClaw auth/runtime state. This PR-body refresh is exact-head proof for the config/default/migration changes and accurately records the unavailable live lanes.

Implementation notes

• Adds src/plugins/slot-resolution.ts as the shared role-slot helper layer for canonical memory role keys, defaults, normalization, effective per-agent resolution, and configured selection collection.
• Updates config schema/types/help/labels/docs for global and per-agent memory role slots.
• Keeps PluginSlotsConfig.contextEngine for global plugin slots, and adds an agent-slot type/schema that omits contextEngine.
• Updates doctor legacy memory-slot handling so doctor --fix canonicalizes to memory.recall while retaining legacy memory during the compatibility window.
• Updates plugin loader/activation/startup resolution so explicit role-slot selections activate memory plugins without confusing default memory-core fallback with a user-selected plugin under restrictive allowlists.
• Extends validation to report missing role-slot plugin refs and validate schemas for plugins selected only by non-recall or per-agent memory roles.
• Extends skill publishing, status scanning, uninstall cleanup, memory runtime, dreaming, compaction provider, and SDK alias behavior to use the role-aware model.
• Keeps memory runtime registrations addressable by plugin id, so runtime recall dispatch follows the resolved global or per-agent memory.recall owner.
• Passes the session agent id into Pi embedded extension factory construction so safeguard compaction provider selection honors per-agent memory.compaction overrides in both normal session runs and manual/queued compaction paths.
• Updates bundled memory plugin manifests/docs to describe role support and the recall shorthand migration.

Compatibility and risk

Backward compatibility preserved:

• Existing plugins.slots.memory configs continue to resolve as recall selection.
• doctor --fix now keeps the legacy plugins.slots.memory key in place after copying to memory.recall, so third-party plugins that still read only the legacy key do not hit an immediate flag day.
• The default recall plugin remains memory-core unless explicitly disabled or overridden.
• plugins.enabled = false, denylist behavior, and explicit plugins.entries.<id>.enabled = false remain respected.
• Global plugins.slots.contextEngine remains separate from memory roles and continues to select context engines such as LosslessClaw.

Compatibility changes reviewers should inspect:

• memory.recall, memory.compaction, memory.capture, memory.dreaming, and memory.userModel become durable public config keys after merge.
• Per-agent plugin-slot overrides are memory-only. agents.list[*].plugins.slots.contextEngine is intentionally rejected until a future PR defines and implements agent-scoped context-engine semantics.
• Legacy plugins.slots.memory is still deprecated as a canonical authoring target even though it remains supported and retained for compatibility.

Primary risks:

• The public memory role taxonomy becomes API surface after merge, so naming and docs need maintainer signoff.
• Role-gating mistakes could attach the wrong memory provider or suppress expected memory behavior; this PR adds focused coverage across loader, validation, startup, runtime, doctor, status, skill publishing, and cleanup surfaces.
• Third-party plugins should eventually read memory.recall, but the retained legacy key avoids forcing that ecosystem update immediately.

Verification

Current pushed head:

• 23a29722524c982f9ed410610395e1db0363857e (fix(memory): polish recall slot defaults)

Exact-head proof-host artifact:

• [REDACTED]/proof-refresh-23a2972252-20260528T172630Z
• Host/runtime: [REDACTED], Node v22.22.2, pnpm 11.2.2, OpenClaw 2026.5.28 (23a2972).
• command-summary.json: deterministicOk=true, deterministicFailures=[].

Exact-head local proof gates:

• corepack pnpm install --frozen-lockfile: passed.
• corepack pnpm build:strict-smoke: passed.
• Focused memory/config/Hermes suite: passed, 10 files / 496 tests.
• Agentic shard: passed, 3 Vitest shards, including 76 files / 1260 tests in the agent shard.
• Plugin memory suite: passed, 5 files / 320 tests.
• Command/config memory suite: passed, 1 file / 14 tests.
• Active-memory extension suite: passed, 1 file / 126 tests.
• Memory-wiki bridge suite: passed, 1 file / 11 tests.
• corepack pnpm tsgo:prod: passed.
• git diff --check origin/main...HEAD: passed.
• Legacy memory doctor compatibility harness: passed; doctor --fix copied memory to memory.recall and retained the legacy key as designed.
• Context-engine/LosslessClaw config-compat harness: passed config validation for plugins.slots.contextEngine = "lossless-claw"; installed runtime execution was unavailable because the plugin was not installed on the proof host.
• node openclaw.mjs models status --probe --json: completed and documented the host's current missing live auth state.

Current GitHub Actions signal for pushed head 23a29722524c982f9ed410610395e1db0363857e at body refresh time:

• Real behavior proof: passed.
• check-lint: passed.
• check-prod-types: passed.
• check-docs: passed.
• Security High (network-ssrf-boundary): passed.
• Security High (core-auth-secrets): passed.
• Critical quality shards shown by gh pr checks are green except Critical Quality (core-auth-secrets), which GitHub reports as failed/cancelled after 25m5s with no failed-step log or assertion output in gh run view.

Live proof limitation:

• The previous body included live GitHub Copilot and Honcho excerpts from stale head bb771d08911d361241d402bf684ca41645623f02. Those excerpts are no longer claimed as current-head proof.
• Exact-head live probes on the available proof host did not have the expected live auth/runtime state: github-copilot auth was missing for the main agent, the GitHub Copilot live harness skipped due to no API keys, lossless-claw was not installed in the plugin registry, and local Honcho health did not respond.