SuperLocalMemory and Bounded Loops are separate optional products. SLM owns durable, profile-scoped memory and learning evidence. Bounded Loops owns graph execution, gate decisions, and receipt truth. Neither package imports the other, and installing SLM never installs, starts, or configures Bounded Loops.
When both are installed, an MCP client can call
observe_bounded_loop_evidence(workspace) to take one explicit read-only
snapshot of terminal graph runs in that workspace. SLM resolves only the
installed bounded-loops-mcp executable; callers cannot supply a shell command
or arguments. The workspace must be an existing absolute path.
The bridge negotiates the producer's advertised MCP capability, not a package version. It requires this exact contract:
bounded-loops.dev/slm-bridge/v1
The producer must advertise bl_graph_evidence with the
observe_terminal_run operation. SLM then lists terminal runs and fetches each
one with its run_ref address. It never sends run_id as a replacement for
that address. An absent or incompatible producer returns a structured refusal;
it does not change SLM memory behaviour.
Compatible evidence is stored in the additive learning.db migration M041,
under the active SLM profile. It preserves the producer's terminal state,
demonstration flag, receipt head, digests, and node gate outcomes. The Living
Brain shows the resulting count as Bounded Loop observations.
This is deliberately observation-only. In v1, eligible_for_learning must be
false, and SLM does not use these rows for recall, ranking, routing, reward, or
automatic behavioural change. A successful graph run is not proof that a
memory decision should be learned. A future contract can add an explicitly
audited learning authority without changing this v1 boundary.
No graph artifacts, commands, paths, or gate prose are copied into SLM. The bridge reads only the producer's public sanitized evidence document. Evidence is deleted with its SLM profile and a durable profile-closure tombstone blocks stale processes from writing it again after erasure.
Observation is an explicit MCP operation and has a five-second producer timeout. Recall and remember never launch Bounded Loops or open this evidence store; they never enter its recall or ranking logic. Observation writes use the bounded learning-db receipt gate and can return a retryable refusal under contention rather than delaying a memory answer.