KuraPre-release

Context & Session

For a personal agent, context management is session management — and in Kura both are plugins on the chat/pre-dispatch waterfall, so any plugin (builtin or external) can modify or replace the policy.

The default context plugin

Runs first in the waterfall. Deterministic, budgeted, always cited:

  1. Memory bootstrap — the tenant’s Ready L3 personas, then L2 scenarios (newest first), injected as system-frame messages under memoryBudgetChars (default 4000). Every injected line carries its citation inline: Memory[l3 mem_xxx] persona: … — recalled memory is evidence, never bare text. L1 atoms are never bulk-injected.
  2. Query-time recall — the turn’s last user message recalls Ready L1 atoms through a three-ranker fusion: BM25 + recency + character-trigram vectors, combined with reciprocal-rank fusion (RRF). Atoms with no lexical or vector affinity are never recalled. Budget: retrievalBudgetChars (default 2000). Injected as Memory[l1 …] (recalled): ….
  3. Symbolic compression — oversized non-frame messages (default >8000 chars) externalize to a full-content L0 memory ref; the window keeps a 200-char preview plus the citation. Token cost drops; the evidence path stays (GET /v1/memory/assets/{id}).
  4. Binding-aware loadout — agent-visibility assets inject only for the active agent profile bound to them.

Every pass emits a context.assembled event carrying the AssemblyRecord: each inclusion (asset, layer, chars, source) and each exclusion with its reason (over_budget/empty_content/visibility). Nothing enters or misses the context silently.

The session-strategy plugin

Runs after context. Frame-preserving window shaping:

  • System messages (persona, skills, safety, injected memory) are the frame — never elided.
  • The most recent keepRecent messages always survive.
  • Over budget, the oldest history collapses into one marker — and the elided span is captured to memory first (thread-linked L0), so the marker cites it: …elided span captured as Memory[l0_ref mem_x]. The model can drill back. Eviction never destroys context.

Two budgets, keyed by source: personal long sessions (48k chars) and IM threads (sourceKind: channel, 16k) — one context per thread.

On-demand retrieval API

The same fused ranking, exposed for clients and tools:

curl -s http://127.0.0.1:19192/v1/retrieval/queries \
  -H 'content-type: application/json' \
  -d '{"query": "which package manager do we use", "limit": 5}'

Hits carry rank, source links, and drill-down member ids.

Last updated: