Team knowledge, findable at session start
Your team has shipped real capabilities any Amplifier session could reuse — but that work lives in individual heads, so each new session rebuilds solved problems from scratch.
The fix isn't a better memory. It's making that prior work findable at the moment you'd reuse it.
"A shared team knowledge base for Amplifier — every session starts aware of what the team has built and searches existing capabilities before building anything new."
README.md line 3 · microsoft/amplifier-bundle-team-knowledge-base
The team-knowledge-base bundle isn't a sketch. It's a shipped Microsoft Amplifier bundle with ~3.3 months of active development and heavy test coverage.
So what exactly does it promise to do? Its thesis is written into the README.
The bundle's core promise is stated up front: every session starts aware of what the team has built, and searches existing capabilities before building anything new.
Two mechanisms deliver that promise. First: awareness, handed to you automatically.
"every session starts aware of what the team has built and searches existing capabilities before building anything new"
README.md line 3 · head -5 README.md
A lifecycle hook fires at session start and injects a summary of team work — and it's on by default, so awareness happens without you remembering to ask.
Awareness alone isn't enough. The next mechanism makes searching non-optional.
A PROACTIVE rule is injected as behavioral text at session start: before building, designing, or brainstorming anything, the agent must call team_knowledge(search) first. You can't skip it.
But an enforced search only helps if the search actually finds the right thing.
hooks-team-knowledge-context/__init__.py lines 83–91
Search runs on all-MiniLM-L6-v2 embeddings over a sqlite-vec index, ranked by embedding distance and merged across every configured team — so one query finds meaning, not just literal keyword hits.
That's why a single enforced search is enough to surface real reuse.
WHERE v.embedding MATCH ? (repo_reader.py 235)ORDER BY v.distance (repo_reader.py 237)Because search is enforced and semantic, a single query surfaces existing pieces owned by different teammates — then layered discovery walks from a capability to its owner to more capabilities. Build-from-scratch becomes compose-from-existing.
The pattern generalizes past this one bundle.
context/instructions-tool.md lines 213–273 (Layered Discovery)
Surface team work automatically and require a search first — so knowledge stops living only in heads, and every session composes on what the team already built.
Data as of: latest local commit 2026-06-28
Feature status: Active — bundle v0.4.0, behavior v0.3.0
Repo: microsoft/amplifier-bundle-team-knowledge-base
Primary contributor: Samuel Lee (59 commits); also Sam Schillace, Sam, Salil Das
Gaps: "23 PRs" is a lower bound from local commit subjects (not queried via gh); the "three teammates" framing is illustrative, not a counted metric.