Lambo: A Living Topological Memory Substrate for Multi-Agent Software Development

Autonomous coding agents operate within ephemeral process lifecycles. When agents complete execution, their intermediate reasoning and architectural context vanish. Similarity retrieval alone does not explicitly encode dependency or invalidation relationships. We present Lambo, an in-memory topological memory daemon for multi-agent software engineering. Lambo models development state as a directed typed graph with an earned canonization lifecycle. Concurrent agents submit asynchronous writes through per-agent lanes with individual receipt states. Context retrieval merges lexical, recency, and dense vector candidates, expands prioritized breadth-first over dependency edges under invalidation semantics, then applies canonical-first ranking and hot-list conflict preservation under token budgets. We report telemetry from two production rigs at nominal 5-minute health sampling. The extract contains 2,495 CUDA and 3,429 Metal heartbeat snapshots. Those are 8.7 and 11.9 interval-equivalent days within 12.7 and 16.0 day windows. Deduplication rates reached 12.0% during synchronized review swarms on Metal (against 0.8% in single-agent sessions) and 4.3% among swarm-named agents on CUDA (against 1.7% in non-swarm streams, 2.2% aggregate). The tracked infrastructure counters were zero in this extract, which does not establish zero failures or data loss. CUDA inspect misses were 22 of 82, while Metal had one miss in four calls and is too small for a cross-rig rate comparison. No concepts reached Canonical status, and the observed Metal store had no concepts eligible for the first promotion gate. A preliminary paired comparison against Claude Code file memory scored 10/12 for file memory against 8/12 for Lambo on one corpus, and 8/12 against 2/12 on a second. We discuss citation dilution as a separate hypothesis and identify controlled retrieval evaluation as future work.