feat(memory): add holographic fact store and MCP tools

Cargo.toml

@@ -120,6 +120,7 @@ crossterm = "0.28"
 fs2 = "0.4"
 sysinfo = { version = "0.32", default-features = false, features = ["system"] }
 tree-sitter-hlsl = { version = "0.2.0", optional = true }
+amari-holographic = "0.23.0"
 
 [target.'cfg(not(windows))'.dependencies]
 flate2 = "1"

src/accounting/parser.rs

@@ -163,7 +163,7 @@ fn parse_line(line: &str, project_hash: &str, session_id: &str) -> Option<CostTu
 }
 
 /// Parse an ISO 8601 timestamp to unix epoch seconds.
-fn parse_timestamp(ts: &str) -> Option<u64> {
+pub(crate) fn parse_timestamp(ts: &str) -> Option<u64> {
     // Handle "2026-04-14T10:32:15.039Z" format
     // Simple parsing without pulling in chrono: split on known positions
     if ts.len() < 19 {

src/agents/claude.rs

@@ -355,8 +355,14 @@ fn install_claude_md_rules(claude_md_path: &Path) -> Result<()> {
         User instructions take precedence over skills.\n\
         - If a code analysis question cannot be fully answered by tokensave MCP tools, \
         try querying the SQLite database directly at `.tokensave/tokensave.db` \
-        (tables: `nodes`, `edges`, `files`). Use SQL to answer complex structural queries \
+        (tables: `nodes`, `edges`, `files`, `memory_facts`, `memory_entities`, \
+        `memory_feedback_events`). Use SQL to answer complex structural queries \
         that go beyond what the built-in tools expose.\n\
+        - For durable project/user facts, prefer `tokensave_fact_store`, \
+        `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. \
+        Use `tokensave_message_search` for project-local Cursor transcript recall when \
+        prior conversation context matters. Do not store secrets, credentials, or \
+        unnecessary PII in persistent facts.\n\
         - If you discover a gap where an extractor, schema, or tokensave tool could be \
         improved to answer a question natively, propose to the user that they open an issue \
         at https://github.com/aovestdipaperino/tokensave describing the limitation. \

src/agents/codex.rs

@@ -244,8 +244,14 @@ fn install_prompt_rules(agents_md: &Path) -> Result<()> {
         faster than file reads.\n\n\
         If a code analysis question cannot be fully answered by tokensave MCP tools, \
         try querying the SQLite database directly at `.tokensave/tokensave.db` \
-        (tables: `nodes`, `edges`, `files`). Use SQL to answer complex structural queries \
+        (tables: `nodes`, `edges`, `files`, `memory_facts`, `memory_entities`, \
+        `memory_feedback_events`). Use SQL to answer complex structural queries \
         that go beyond what the built-in tools expose.\n\n\
+        For durable project/user facts, prefer `tokensave_fact_store`, \
+        `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. \
+        Use `tokensave_message_search` for project-local Cursor transcript recall when \
+        prior conversation context matters. Do not store secrets, credentials, or \
+        unnecessary PII in persistent facts.\n\n\
         If you discover a gap where an extractor, schema, or tokensave tool could be \
         improved to answer a question natively, propose to the user that they open an issue \
         at https://github.com/aovestdipaperino/tokensave describing the limitation. \

src/agents/copilot.rs

@@ -399,7 +399,9 @@ fn install_prompt_rules(instructions_path: &Path) -> Result<()> {
         that go beyond what the built-in tools expose.\n\n\
         For durable project/user facts, prefer `tokensave_fact_store`, \
         `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. \
-        Do not store secrets, credentials, or unnecessary PII in persistent facts.\n\n\
+        Use `tokensave_message_search` for project-local Cursor transcript recall when \
+        prior conversation context matters. Do not store secrets, credentials, or \
+        unnecessary PII in persistent facts.\n\n\
         If you find a gap where tokensave could answer a question natively, propose opening \
         an issue at https://github.com/aovestdipaperino/tokensave. Remind the user to strip \
         sensitive or proprietary code from any issue text before submitting.\n"

src/agents/cursor.rs

@@ -161,6 +161,7 @@ alwaysApply: true
 
 - For codebase exploration, symbol lookup, call graphs, callers/callees, impact analysis, affected files, and architectural navigation, use the tokensave MCP tools first.
 - Prefer tools such as `tokensave_context`, `tokensave_search`, `tokensave_callers`, `tokensave_callees`, `tokensave_impact`, `tokensave_files`, `tokensave_affected`, and related read-only tokensave tools before broad file reads or search.
+- For durable project/user facts, prefer `tokensave_fact_store`, `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. Use `tokensave_message_search` for project-local Cursor transcript recall when prior conversation context matters.
 - Only fall back to regular file reads, search, or shell commands when tokensave cannot answer the question or after tokensave has identified the exact files or symbols to inspect.
 "#;
     write_generated_text(rule_path, contents)?;
@@ -284,6 +285,17 @@ fn install_hooks(hooks_path: &Path, tokensave_bin: &str) -> Result<()> {
         60,
         None,
     );
+    // End-of-turn transcript ingestion. This is the primary, off-hot-path place
+    // we capture Cursor transcripts (beforeSubmitPrompt only does a tiny tail
+    // read), so it gets a generous timeout for the incremental catch-up.
+    install_cursor_hook_entry(
+        &mut hooks,
+        "stop",
+        tokensave_bin,
+        "hook-cursor-stop",
+        30,
+        None,
+    );
 
     safe_write_json_file(hooks_path, &hooks, backup.as_deref())?;
     eprintln!(
@@ -502,6 +514,7 @@ fn doctor_check_hooks(dc: &mut DoctorCounters, hooks_path: &Path) {
         ("afterFileEdit", "hook-cursor-after-file-edit"),
         ("afterShellExecution", "hook-cursor-after-shell"),
         ("workspaceOpen", "hook-cursor-workspace-open"),
+        ("stop", "hook-cursor-stop"),
     ];
     let missing: Vec<&str> = expected
         .iter()

src/agents/gemini.rs

@@ -158,8 +158,14 @@ fn install_prompt_rules(gemini_md: &Path) -> Result<()> {
         faster than file reads.\n\n\
         If a code analysis question cannot be fully answered by tokensave MCP tools, \
         try querying the SQLite database directly at `.tokensave/tokensave.db` \
-        (tables: `nodes`, `edges`, `files`). Use SQL to answer complex structural queries \
+        (tables: `nodes`, `edges`, `files`, `memory_facts`, `memory_entities`, \
+        `memory_feedback_events`). Use SQL to answer complex structural queries \
         that go beyond what the built-in tools expose.\n\n\
+        For durable project/user facts, prefer `tokensave_fact_store`, \
+        `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. \
+        Use `tokensave_message_search` for project-local Cursor transcript recall when \
+        prior conversation context matters. Do not store secrets, credentials, or \
+        unnecessary PII in persistent facts.\n\n\
         If you discover a gap where an extractor, schema, or tokensave tool could be \
         improved to answer a question natively, propose to the user that they open an issue \
         at https://github.com/aovestdipaperino/tokensave describing the limitation. \

src/agents/kimi.rs

@@ -159,8 +159,14 @@ fn install_prompt_rules(agents_md: &Path) -> Result<()> {
         faster than file reads.\n\n\
         If a code analysis question cannot be fully answered by tokensave MCP tools, \
         try querying the SQLite database directly at `.tokensave/tokensave.db` \
-        (tables: `nodes`, `edges`, `files`). Use SQL to answer complex structural queries \
+        (tables: `nodes`, `edges`, `files`, `memory_facts`, `memory_entities`, \
+        `memory_feedback_events`). Use SQL to answer complex structural queries \
         that go beyond what the built-in tools expose.\n\n\
+        For durable project/user facts, prefer `tokensave_fact_store`, \
+        `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. \
+        Use `tokensave_message_search` for project-local Cursor transcript recall when \
+        prior conversation context matters. Do not store secrets, credentials, or \
+        unnecessary PII in persistent facts.\n\n\
         If you discover a gap where an extractor, schema, or tokensave tool could be \
         improved to answer a question natively, propose to the user that they open an issue \
         at https://github.com/aovestdipaperino/tokensave describing the limitation. \

src/agents/kiro.rs

@@ -449,7 +449,13 @@ have been tried. Delegation is still appropriate for long-running execution work
 such as builds, tests, generated reports, or independent implementation tasks.\n\n\
 If a code analysis question cannot be fully answered by tokensave MCP tools, try \
 querying the SQLite database directly at `.tokensave/tokensave.db` (tables: `nodes`, \
-`edges`, `files`). Use SQL for structural queries that go beyond the MCP tools.\n\n\
+`edges`, `files`, `memory_facts`, `memory_entities`, `memory_feedback_events`). \
+Use SQL for structural queries that go beyond the MCP tools.\n\n\
+For durable project/user facts, prefer `tokensave_fact_store`, \
+`tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. Use \
+`tokensave_message_search` for project-local Cursor transcript recall when prior \
+conversation context matters. Do not store secrets, credentials, or unnecessary PII \
+in persistent facts.\n\n\
 If you discover a gap where an extractor, schema, or tokensave tool could answer a \
 question natively, propose opening an issue at \
 https://github.com/aovestdipaperino/tokensave. Remind the user to strip sensitive \

src/agents/opencode.rs

@@ -191,8 +191,14 @@ fn install_prompt_rules(prompt_path: &Path) -> Result<()> {
         faster than file reads.\n\n\
         If a code analysis question cannot be fully answered by tokensave MCP tools, \
         try querying the SQLite database directly at `.tokensave/tokensave.db` \
-        (tables: `nodes`, `edges`, `files`). Use SQL to answer complex structural queries \
+        (tables: `nodes`, `edges`, `files`, `memory_facts`, `memory_entities`, \
+        `memory_feedback_events`). Use SQL to answer complex structural queries \
         that go beyond what the built-in tools expose.\n\n\
+        For durable project/user facts, prefer `tokensave_fact_store`, \
+        `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. \
+        Use `tokensave_message_search` for project-local Cursor transcript recall when \
+        prior conversation context matters. Do not store secrets, credentials, or \
+        unnecessary PII in persistent facts.\n\n\
         If you discover a gap where an extractor, schema, or tokensave tool could be \
         improved to answer a question natively, propose to the user that they open an issue \
         at https://github.com/aovestdipaperino/tokensave describing the limitation. \

src/agents/vibe.rs

@@ -191,7 +191,9 @@ fn install_prompt_rules(prompt_path: &Path) -> Result<()> {
         that go beyond what the built-in tools expose.\n\n\
         For durable project/user facts, prefer `tokensave_fact_store`, \
         `tokensave_fact_feedback`, and `tokensave_memory_status` over ad-hoc notes. \
-        Do not store secrets, credentials, or unnecessary PII in persistent facts.\n\n\
+        Use `tokensave_message_search` for project-local Cursor transcript recall when \
+        prior conversation context matters. Do not store secrets, credentials, or \
+        unnecessary PII in persistent facts.\n\n\
         If you find a gap where tokensave could answer a question natively, propose opening \
         an issue at https://github.com/aovestdipaperino/tokensave. Remind the user to strip \
         sensitive or proprietary code from any issue text before submitting.\n\n\

src/cli.rs

@@ -143,6 +143,9 @@ pub enum Commands {
     /// Cursor workspaceOpen hook handler (called by Cursor, not by users directly)
     #[command(name = "hook-cursor-workspace-open", hide = true)]
     HookCursorWorkspaceOpen,
+    /// Cursor stop hook handler (called by Cursor, not by users directly)
+    #[command(name = "hook-cursor-stop", hide = true)]
+    HookCursorStop,
     /// Codex SessionStart hook handler (called by Codex, not by users directly)
     #[command(name = "hook-codex-session-start", hide = true)]
     HookCodexSessionStart,