Psychic Cache
The operational memory layer: a Zod-validated type system with a closed tag taxonomy.
The psychic cache is the operational memory layer — fast, structured, and queryable. Every entry is validated at the gateway against a discriminated union keyed on context_type.
Type system
| Type | Writer | Purpose | Key fields |
|---|---|---|---|
decision | LISA | Resolved choices | decision_rationale, decision_alternatives[] |
directive | LISA | Task briefs and updates | directive_target, directive_priority |
intel | LISA / Agent | Findings, facts | intel_source, intel_confidence |
state_change | LISA | Status transitions | state_from, state_to, state_trigger |
output | Agent | Deliverable links | agent_name, output_url, output_summary |
report | LISA | Post-dispatch meta-observations | report_dispatch_id, report_outcome |
escalation | Agent | Re-routing signals | escalation_target, escalation_reason |
shadow_clone | Agent | Clone requests | subtasks[], clone_count |
thread_checkpoint | Hook script | Pre-compaction state | session_id, turn_count, working_state_prose |
A missing required field fails the write with a legible error. Extra unknown fields are rejected. The structured field — not free text stuffed into content — is what makes an entry queryable and auditable.
Tag taxonomy (closed set)
- Domain tags (8) —
security,verification,engineering,planning,creative,vault,legal,data. - Topic tags (7) —
pricing,branding,stakeholder,timeline,risk,scope,deliverable.
Every entry carries at least one domain tag and up to two topic tags. Inside a dispatch, every write also carries the dispatch_id so the cache is attributable to the work that produced it.
The raw-write path
Structured writes go through a raw payload path — a single JSON string field validated server-side — so that mid-session schema evolution never breaks a client whose tool schema was cached at startup. See Raw twin discipline.