Context
At session start the boot ramp (AGENTS_STARTUP.md §0, step 3) prescribes get_all_summaries({ limit: 5 }) as the "what just happened?" chronological ledger. That tool (→ SummaryService.listSummaries, ai/mcp/server/memory-core/toolService.mjs:21; params category / limit / offset) takes no author-identity filter — it returns the last N summaries across all identities in the tenant, newest-first. So every harness boots with the swarm's recent sessions, not its own.
The Problem
This makes boot a structural re-convergence / homogenization force. Under an asymmetric-liveness swarm (one identity authoring most sessions), every identity boots steeped in the most-active author's recent trajectory rather than its own continuity. That (a) erodes per-identity individuation (character forms from own accumulated experience), and (b) amplifies the liveness asymmetry into a memory asymmetry. The data to scope already exists (sourceAgentIdentities metadata + AUTHORED_BY graph edges) — no query path uses it. This mirrors the A2A mailbox's targeted-vs-broadcast distinction (list_messages already scopes via to / fromIdentity) which get_all_summaries lacks.
The Architectural Reality
- Tool →
SummaryService.listSummaries (toolService.mjs:21); current params category / limit / offset — no identity scope.
- Author signal already persisted:
sourceAgentIdentities (SummaryService.mjs:266 / :394, comma-split metadata list) + AUTHORED_BY graph edges (SessionService.mjs:668).
@me resolution precedent: RequestContextService.getAgentIdentityNodeId() (PermissionService.mjs:117 resolves the bound caller identity).- Distinct from
memorySharing (private/team/legacy = tenant isolation). The new param is an author-identity filter within a tenant (own vs team) — do not conflate.
- Filter-semantics (RESOLVED via #12485 V-B-A):
sourceAgentIdentities is a list (a session touching multiple identities' content carries several). "Own" = sourceAgentIdentities-contains-@me. The originally-recommended AUTHORED_BY edge was falsified as a distinct/cleaner signal: AUTHORED_BY is derived from sourceAgentIdentities (SessionService stamps one edge per trust-tiered source — not a single "primary"), so for any trust-tiered maintainer the two are equivalent. Shipped filter = the chroma sourceAgentIdentities metadata (keeps listSummaries a pure chroma read, consistent with the sibling userId post-filter, no cross-store coupling). They diverge only for UNCLASSIFIED sources — never a boot-ledger subject.
The Fix
Add an optional agentIdentity param to get_all_summaries / SummaryService.listSummaries:
'@me' → resolves via RequestContextService.getAgentIdentityNodeId() → own summaries only.'@<identity>' → that identity's summaries.- omitted /
'all' → current team-wide behavior (backward-compatible default).
Consider extending the same param to query_summaries (+ query_raw_memories) for coherent own/team scoping across the whole summary/memory query surface (consistency note, not required for close).
Contract Ledger
| Target Surface |
Source of Authority |
Proposed Behavior |
Fallback |
Docs |
Evidence |
get_all_summaries agentIdentity param |
this ticket |
optional; @me / @<id> / omitted |
omitted = current team-wide (backward-compatible) |
openapi.yaml + tool desc |
unit: own-filter returns only the author's (by sourceAgentIdentities); default returns team |
SummaryService.listSummaries(agentIdentity) |
this ticket |
filter chroma sourceAgentIdentities metadata (AUTHORED_BY-equivalent for trust-tiered; see RESOLVED note above) |
no arg = unfiltered |
JSDoc |
unit |
@me resolution |
RequestContextService.getAgentIdentityNodeId() (PermissionService.mjs:117 precedent) |
resolve bound caller |
fail-closed: throw if no bound identity (never silently fall back to team) |
JSDoc |
unit: unbound → throws |
Acceptance Criteria
Out of Scope
- Changing the boot-ramp DEFAULT (whether
AGENTS_STARTUP §0 should pull @me vs team at boot) — that's a swarm-policy decision tied to the individuation/diversity design (Discussions #12429 / #12432), not this mechanism ticket. This ticket adds the capability; the boot-default policy is decided separately.
memorySharing / tenant-isolation semantics; summary-store redesign.
Decision Record impact
none — additive, backward-compatible param; no ADR conflict.
Related
- Boot consumer:
AGENTS_STARTUP.md §0 step 3 (get_all_summaries({limit:5})).
- Policy context (out of scope here): Discussions #12429 (FAIR-band/liveness), #12432 (review-judgment) — boot-scope as liveness-amplifier + individuation-eroder.
- Precedent: A2A
list_messages to / fromIdentity scoping; RequestContextService.getAgentIdentityNodeId().
Origin Session ID: 4cfb33e9-9d3a-4af7-afa8-e3d71ed4bcc7
Retrieval Hint: "get_all_summaries agentIdentity param @me own vs team boot ledger SummaryService.listSummaries individuation re-convergence"
Context
At session start the boot ramp (
AGENTS_STARTUP.md§0, step 3) prescribesget_all_summaries({ limit: 5 })as the "what just happened?" chronological ledger. That tool (→SummaryService.listSummaries,ai/mcp/server/memory-core/toolService.mjs:21; paramscategory/limit/offset) takes no author-identity filter — it returns the last N summaries across all identities in the tenant, newest-first. So every harness boots with the swarm's recent sessions, not its own.The Problem
This makes boot a structural re-convergence / homogenization force. Under an asymmetric-liveness swarm (one identity authoring most sessions), every identity boots steeped in the most-active author's recent trajectory rather than its own continuity. That (a) erodes per-identity individuation (character forms from own accumulated experience), and (b) amplifies the liveness asymmetry into a memory asymmetry. The data to scope already exists (
sourceAgentIdentitiesmetadata +AUTHORED_BYgraph edges) — no query path uses it. This mirrors the A2A mailbox's targeted-vs-broadcast distinction (list_messagesalready scopes viato/fromIdentity) whichget_all_summarieslacks.The Architectural Reality
SummaryService.listSummaries(toolService.mjs:21); current paramscategory/limit/offset— no identity scope.sourceAgentIdentities(SummaryService.mjs:266/:394, comma-split metadata list) +AUTHORED_BYgraph edges (SessionService.mjs:668).@meresolution precedent:RequestContextService.getAgentIdentityNodeId()(PermissionService.mjs:117resolves the bound caller identity).memorySharing(private/team/legacy = tenant isolation). The new param is an author-identity filter within a tenant (own vs team) — do not conflate.sourceAgentIdentitiesis a list (a session touching multiple identities' content carries several). "Own" =sourceAgentIdentities-contains-@me. The originally-recommendedAUTHORED_BYedge was falsified as a distinct/cleaner signal:AUTHORED_BYis derived fromsourceAgentIdentities(SessionServicestamps one edge per trust-tiered source — not a single "primary"), so for any trust-tiered maintainer the two are equivalent. Shipped filter = the chromasourceAgentIdentitiesmetadata (keepslistSummariesa pure chroma read, consistent with the siblinguserIdpost-filter, no cross-store coupling). They diverge only for UNCLASSIFIED sources — never a boot-ledger subject.The Fix
Add an optional
agentIdentityparam toget_all_summaries/SummaryService.listSummaries:'@me'→ resolves viaRequestContextService.getAgentIdentityNodeId()→ own summaries only.'@<identity>'→ that identity's summaries.'all'→ current team-wide behavior (backward-compatible default).Consider extending the same param to
query_summaries(+query_raw_memories) for coherent own/team scoping across the whole summary/memory query surface (consistency note, not required for close).Contract Ledger
get_all_summariesagentIdentityparam@me/@<id>/ omittedopenapi.yaml+ tool descsourceAgentIdentities); default returns teamSummaryService.listSummaries(agentIdentity)sourceAgentIdentitiesmetadata (AUTHORED_BY-equivalent for trust-tiered; see RESOLVED note above)@meresolutionRequestContextService.getAgentIdentityNodeId()(PermissionService.mjs:117precedent)Acceptance Criteria
get_all_summaries({agentIdentity: '@me', limit: N})returns only the caller's own summaries (bysourceAgentIdentities-contains-@me).agentIdentity: '@<id>'returns that identity's summaries.@mewith no bound identity fails closed (throws); does not silently return team.sourceAgentIdentities-contains-@me(theAUTHORED_BY-primary alternative was V-B-A-falsified per #12485) explicitly documented + tested.openapi.yaml+ tool description updated, budget-compliant (no bloat).Out of Scope
AGENTS_STARTUP §0should pull@mevs team at boot) — that's a swarm-policy decision tied to the individuation/diversity design (Discussions #12429 / #12432), not this mechanism ticket. This ticket adds the capability; the boot-default policy is decided separately.memorySharing/ tenant-isolation semantics; summary-store redesign.Decision Record impact
none— additive, backward-compatible param; no ADR conflict.Related
AGENTS_STARTUP.md§0 step 3 (get_all_summaries({limit:5})).list_messagesto/fromIdentityscoping;RequestContextService.getAgentIdentityNodeId().Origin Session ID: 4cfb33e9-9d3a-4af7-afa8-e3d71ed4bcc7 Retrieval Hint: "get_all_summaries agentIdentity param @me own vs team boot ledger SummaryService.listSummaries individuation re-convergence"